Merge branch 'master' of https://github.com/biometrics/openbr

bklare
2 parents 50f1dd39 c73a5623
Showing 48 changed files with 635 additions and 560 deletions
3rdparty/stasm/stasm/CMakeLists.txt
3rdparty/stasm/stasm/stasm/src/print.cpp
3rdparty/stasm4.0.0/stasm/MOD_1/facedet.cpp
3rdparty/stasm4.0.0/stasm/MOD_1/facedet.h
3rdparty/stasm4.0.0/stasm/stasm_lib.cpp
3rdparty/stasm4.0.0/stasm/stasm_lib.h
CMakeLists.txt
app/br/br.cpp
docs/docs/api_docs/cl_api.md
docs/docs/install.md
openbr/core/boost.cpp
openbr/core/boost.h
openbr/core/eval.cpp
openbr/core/eval.h
openbr/core/opencvutils.cpp
openbr/core/opencvutils.h
openbr/core/plot.cpp
openbr/core/plot.h
openbr/openbr.cpp
openbr/openbr.h
-# Stasm CMakeLists
-
-aux_source_directory(stasm/src STASM_SRC)
-include_directories(stasm/include)
-
-#aux_source_directory(tasm/src TASM_SRC)
-#include_directories(tasm/include)
-
-add_library(stasm SHARED ${STASM_SRC} ${TASM_SRC})
-qt5_use_modules(stasm ${QT_DEPENDENCIES})
-
-set_target_properties(stasm PROPERTIES
-                                   DEFINE_SYMBOL STASM_LIBRARY
-                                   VERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR}.${CPACK_PACKAGE_VERSION_PATCH}
-                                   SOVERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR}
-                                   LINK_INTERFACE_LIBRARIES ""
-                                   AUTOMOC TRUE)
-
-target_link_libraries(stasm ${OpenCV_LIBS} ${Qt5Core_QTMAIN_LIBRARIES})
-
-file(GLOB STASM_HEADERS stasm/include/*.h)
-install(FILES ${STASM_HEADERS} DESTINATION include/stasm)
-file(GLOB CLASSIC_HEADERS stasm/include/classic/*.mh)
-install(FILES ${CLASSIC_HEADERS} DESTINATION include/stasm/classic)
-file(GLOB HAT_HEADERS stasm/include/hat/*.mh)
-install(FILES ${HAT_HEADERS} DESTINATION include/stasm/hat)
-#file(GLOB TASM_HEADERS tasm/include/*.h)
-#install(FILES ${TASM_HEADERS} DESTINATION include/tasm)
-
-install(TARGETS stasm RUNTIME DESTINATION bin LIBRARY DESTINATION lib ARCHIVE DESTINATION lib)
-// print.cpp: printing and logging utilities for the Stasm library
-//
-// Copyright (C) 2005-2013, Stephen Milborrow
-
-#include "print.h"
-#include "err.h"
-#include "misc.h"
-
-#include <stdio.h>
-#include <sys/stat.h>
-#include <stdarg.h>
-
-namespace stasm
-{
-bool print_g;   // true to allow output to stdout (but error msgs always printed)
-
-bool trace_g;   // true to trace Stasm internal operation
-
-static FILE* logfile_g;  // lprintfs go to this log file as well as stdout
-
-//-----------------------------------------------------------------------------
-
-// Open the log file. After this, when you call lprintf, you print to the log
-// file (as well as to stdout).  This inits the global variable logfile_g.
-
-void OpenLogFile(     // also inits the global variable logfile_g
-    const char* path) // in: log file path, default is "stasm.log"
-{
-    if (!logfile_g)
-    {
-        if (print_g)
-            printf("Generating %s\n", path);
-        logfile_g = fopen(path, "wb");
-        if (!logfile_g)
-            Err("Cannot open \"%s\"", path);
-        // check that we can write to the log file
-        if (fputs("log file\n", logfile_g) < 0)
-            Err("Cannot write to \"%s\"", path);
-        rewind(logfile_g); // rewind so above test msg is not in the log file
-    }
-}
-
-// Like printf but only prints if print_g flag is set.
-// Also prints to the log file if it is open (regardless of print_g).
-
-void lprintf(const char* format, ...)   // args like printf
-{
-    char s[SBIG];
-    va_list args;
-    va_start(args, format);
-    VSPRINTF(s, format, args);
-    va_end(args);
-    if (print_g)
-    {
-        printf("%s", s);
-        fflush(stdout); // flush so if there is a crash we can see what happened
-    }
-    if (logfile_g)
-    {
-        // we don't check fputs here, to prevent recursive calls and msgs
-        fputs(s, logfile_g);
-        fflush(logfile_g);
-    }
-}
-
-// Like printf but prints to the log file only (and not to stdout).
-// Used for detailed stuff that we don't usually want to see.
-
-void logprintf(const char* format, ...) // args like printf
-{
-    if (logfile_g)
-    {
-        char s[SBIG];
-        va_list args;
-        va_start(args, format);
-        VSPRINTF(s, format, args);
-        va_end(args);
-        // we don't check fputs here, to prevent recursive calls and msgs
-        fputs(s, logfile_g);
-        fflush(logfile_g);
-    }
-}
-
-// Like lprintf but always prints even if print_g is false.
-
-void lprintf_always(const char* format, ...)
-{
-    char s[SBIG];
-    va_list args;
-    va_start(args, format);
-    VSPRINTF(s, format, args);
-    va_end(args);
-    printf("%s", s);
-    fflush(stdout); // flush so if there is a crash we can see what happened
-    if (logfile_g)
-    {
-        // we don't check fputs here, to prevent recursive calls and msgs
-        fputs(s, logfile_g);
-        fflush(logfile_g);
-    }
-}
-
-// Like puts but prints to the log file as well if it is open,
-// and does not append a newline.
-
-void lputs(const char* s)
-{
-    printf("%s", s);
-    fflush(stdout);     // flush so if there is a crash we can see what happened
-    logprintf("%s", s);
-}
-
-// Print message only once on the screen, and only 100 times to the log file.
-// This is used when similar messages could be printed many times and it
-// suffices to let the user know just once.  By convention the message is
-// printed followed by "..." so the user knows that just the first message
-// was printed.  The user can look in the log file for further messages if
-// necessary (but we print only 100 times to the log file --- else all the
-// log prints make tasm slow).
-
-void PrintOnce(
-    int&        printed,     // io: zero=print, nonzero=no print
-    const char* format, ...) // in: args like printf
-{
-    char s[SBIG];
-    va_list args;
-    va_start(args, format);
-    VSPRINTF(s, format, args);
-    va_end(args);
-    if (printed == 0 && print_g)
-    {
-        printed = 1;
-        printf("%s", s);
-        fflush(stdout); // flush so if there is a crash we can see what happened
-    }
-    if (printed < 100 && logfile_g)
-    {
-        fputs(s, logfile_g);
-        fflush(logfile_g);
-        printed++;
-        if (printed == 100)
-            logprintf("no more prints of the above message (printed == 100)\n");
-    }
-}
-
-} // namespace stasm
@@ -47,28 +47,16 @@ void DetectFaces(          // all face rects into detpars
     int          minwidth,
     cv::CascadeClassifier cascade) // in: as percent of img width
 {
-    int leftborder = 0, topborder = 0; // border size in pixels
-    Image bordered_img(BORDER_FRAC == 0?
-                       img: EnborderImg(leftborder, topborder, img));
-
-    // Detection results are very slightly better with equalization
-    // (tested on the MUCT images, which are not pre-equalized), and
-    // it's quick enough to equalize (roughly 10ms on a 1.6 GHz laptop).
-
-    Image equalized_img; cv::equalizeHist(bordered_img, equalized_img);
-
     CV_Assert(minwidth >= 1 && minwidth <= 100);
-    int minpix = MAX(100, cvRound(img.cols * minwidth / 100.));
-
     // the params below are accurate but slow
-    static const double SCALE_FACTOR   = 1.1;
-    static const int    MIN_NEIGHBORS  = 3;
+    static const double SCALE_FACTOR   = 1.2;
+    static const int    MIN_NEIGHBORS  = 5;
     static const int    DETECTOR_FLAGS = 0;
     vec_Rect facerects = // all face rects in image
-        Detect(equalized_img, &cascade, NULL,
-               SCALE_FACTOR, MIN_NEIGHBORS, DETECTOR_FLAGS, minpix);
+        Detect(img, &cascade, NULL,
+               SCALE_FACTOR, MIN_NEIGHBORS, DETECTOR_FLAGS, 64);
     // copy face rects into the detpars vector
@@ -80,8 +68,6 @@ void DetectFaces(          // all face rects into detpars
         // detpar.x and detpar.y is the center of the face rectangle
         detpar.x = facerect->x + facerect->width / 2.;
         detpar.y = facerect->y + facerect->height / 2.;
-        detpar.x -= leftborder; // discount the border we added earlier
-        detpar.y -= topborder;
         detpar.width  = double(facerect->width);
         detpar.height = double(facerect->height);
         detpar.yaw = 0; // assume face has no yaw in this version of Stasm
@@ -187,10 +173,10 @@ void FaceDet::DetectFaces_(  // call once per image to find all the faces
     //CV_Assert(!facedet_g.empty()); // check that OpenFaceDetector_ was called
     DetectFaces(detpars_, img, minwidth, cascade);
-    DiscardMissizedFaces(detpars_);
+    //DiscardMissizedFaces(detpars_);
     if (multiface) // order faces on increasing distance from left margin
     {
-        sort(detpars_.begin(), detpars_.end(), IncreasingLeftMargin);
+        sort(detpars_.begin(), detpars_.end(), DecreasingWidth);
     }
     else
     {
@@ -38,13 +38,11 @@ public:
     FaceDet() {}                  // constructor
-
-private:
-    vector<DetPar>  detpars_;     // all the valid faces in the current image
-
     int             iface_;       // index of current face for NextFace_
                                   // indexes into detpars_
+    vector<DetPar>  detpars_;     // all the valid faces in the current image
+private:
     DISALLOW_COPY_AND_ASSIGN(FaceDet);
 }; // end class FaceDet
@@ -160,7 +160,8 @@ int stasm_search_auto_ext( // extended version of stasm_search_auto
     const char* data,
     const int width,
     const int height,
-    StasmCascadeClassifier cascade)
+    StasmCascadeClassifier cascade,
+    FaceDet &detection)
 {
     int returnval = 1;     // assume success
     *foundface = 0;        // but assume no face found
@@ -176,15 +177,14 @@ int stasm_search_auto_ext( // extended version of stasm_search_auto
         // Allocate image
         Image img = Image(height, width,(unsigned char*)data);
-        FaceDet facedet;
-
         // call the face detector to detect the face rectangle(s)
-        facedet.DetectFaces_(img, NULL, false, 10, NULL, cascade.faceCascade);
+        if (detection.detpars_.empty())
+            detection.DetectFaces_(img, NULL, true, 10, NULL, cascade.faceCascade);
         // Get the start shape for the next face in the image, and the ROI around it.
         // The shape will be wrt the ROI frame.
         if (NextStartShapeAndRoi(shape, face_roi, detpar_roi, detpar,
-                                 img, mods_g, facedet, cascade))
+                                 img, mods_g, detection, cascade))
         {
             // now working with maybe flipped ROI and start shape in ROI frame
             *foundface = 1;
@@ -219,9 +219,10 @@ int stasm_search_auto(// call repeatedly to find all faces
     const char *data,
     const int width,
     const int height,
-    StasmCascadeClassifier cascade)
+    StasmCascadeClassifier cascade,
+    FaceDet &detection)
 {
-    return stasm_search_auto_ext(foundface, landmarks, NULL, data, width, height, cascade);
+    return stasm_search_auto_ext(foundface, landmarks, NULL, data, width, height, cascade, detection);
 }
 int stasm_search_single(   // wrapper for stasm_search_auto and friends
@@ -237,7 +238,8 @@ int stasm_search_single(   // wrapper for stasm_search_auto and friends
     (void) datadir;
     (void) imgpath;
-    return stasm_search_auto(foundface, landmarks, img, width, height, cascade);
+    FaceDet detection;
+    return stasm_search_auto(foundface, landmarks, img, width, height, cascade, detection);
 }
 int stasm_search_pinned(    // call after the user has pinned some points
@@ -64,6 +64,10 @@
 #include "stasmcascadeclassifier.h"
+namespace stasm {
+    class FaceDet;
+}
+
 static const int stasm_NLANDMARKS = 77; // number of landmarks
 extern const char* const stasm_VERSION;
@@ -89,7 +93,8 @@ int stasm_search_auto(       // call repeatedly to find all faces
     const char*  data,
     const int    width,
     const int    height,
-    StasmCascadeClassifier cascade);
+    StasmCascadeClassifier cascade,
+    stasm::FaceDet &detection);
 extern "C"
 int stasm_search_single(     // wrapper for stasm_search_auto and friends
@@ -13,7 +13,7 @@ if(NOT DEFINED CPACK_PACKAGE_VERSION_MAJOR)
   set(CPACK_PACKAGE_VERSION_MAJOR 1)
 endif()
 if(NOT DEFINED CPACK_PACKAGE_VERSION_MINOR)
-  set(CPACK_PACKAGE_VERSION_MINOR 0)
+  set(CPACK_PACKAGE_VERSION_MINOR 1)
 endif()
 if(NOT DEFINED CPACK_PACKAGE_VERSION_PATCH)
   set(CPACK_PACKAGE_VERSION_PATCH 0)
@@ -158,8 +158,8 @@ endif()
 # Download the models
 ExternalProject_Add(models
-                    URL http://github.com/biometrics/openbr/releases/download/v1.0.0/models.tar.gz
-                    URL_MD5 0a7c79226d6629954aa32c835a1007b9
+                    URL http://github.com/biometrics/openbr/releases/download/v1.1.0/models.tar.gz
+                    URL_MD5 26cf71abd48cec8d7b2abf3e8f4ebfc4
                     SOURCE_DIR "${PROJECT_SOURCE_DIR}/share/openbr/models"
                     CONFIGURE_COMMAND ""
                     BUILD_COMMAND ""
@@ -169,8 +169,8 @@ public:
                 check(parc >= 2 && parc <= 4, "Incorrect parameter count for 'evalRegression'.");
                 br_eval_regression(parv[0], parv[1], parc >= 3 ? parv[2] : "", parc >= 4 ? parv[3] : "");
             } else if (!strcmp(fun, "evalKNN")) {
-                check(parc == 3, "Incorrect parameter count for 'evalKNN'.");
-                br_eval_knn(parv[0], parv[1], parv[2]);
+                check(parc >= 2 && parc <= 3, "Incorrect parameter count for 'evalKNN'.");
+                br_eval_knn(parv[0], parv[1], parc > 2 ? parv[2] : "");
             } else if (!strcmp(fun, "pairwiseCompare")) {
                 check((parc >= 2) && (parc <= 3), "Incorrect parameter count for 'pairwiseCompare'.");
                 br_pairwise_compare(parv[0], parv[1], parc == 3 ? parv[2] : "");
@@ -186,6 +186,9 @@ public:
             } else if (!strcmp(fun, "plotMetadata")) {
                 check(parc >= 2, "Incorrect parameter count for 'plotMetadata'.");
                 br_plot_metadata(parc-1, parv, parv[parc-1], true);
+            } else if (!strcmp(fun, "plotKNN")) {
+                check(parc >=2, "Incorrect parameter count for 'plotKNN'.");
+                br_plot_knn(parc-1, parv, parv[parc-1], true);
             } else if (!strcmp(fun, "project")) {
                 check(parc == 2, "Insufficient parameter count for 'project'.");
                 br_project(parv[0], parv[1]);
@@ -279,13 +282,14 @@ private:
                "-evalDetection <predicted_gallery> <truth_gallery> [{csv}] [{normalize}] [{minSize}] [{maxSize}]\n"
                "-evalLandmarking <predicted_gallery> <truth_gallery> [{csv} [<normalization_index_a> <normalization_index_b>] [sample_index] [total_examples]]\n"
                "-evalRegression <predicted_gallery> <truth_gallery> <predicted property name> <ground truth property name>\n"
-               "-evalKNN <knn_graph> <knn_truth> [{iet_file}]\n"
+               "-evalKNN <knn_graph> <knn_truth> [{csv}]\n"
                "-pairwiseCompare <target_gallery> <query_gallery> [{output}]\n"
                "-inplaceEval <simmat> <target> <query> [{csv}]\n"
                "-assertEval <simmat> <mask> <accuracy>\n"
                "-plotDetection <file> ... <file> {destination}\n"
                "-plotLandmarking <file> ... <file> {destination}\n"
                "-plotMetadata <file> ... <file> <columns>\n"
+               "-plotKNN <file> ... <file> {destination}\n"
                "-project <input_gallery> {output_gallery}\n"
                "-deduplicate <input_gallery> <output_gallery> <threshold>\n"
                "-likely <input_type> <output_type> <output_likely_source>\n"
@@ -59,6 +59,17 @@ DOCUMENT ME
 * **wraps:** [br_pairwise_compare](c_api/functions.md#br_pairwise_compare)
+### -crossValidate {: #crossvalidate }
+
+Either performs n fold cross validation (if nFolds > 0), or a single iteration of a train / test split with the first of the n folds as the test partition (nFolds < 0).
+
+* **arguments:**
+
+    -crossValidate <nFolds> -algorithm "CrossValidate(description=$ALG,randomSeed=5):CrossValidate+Dist" -train <inputFile> <model>
+
+    -crossValidate <nFolds> -algorithm <model> -enroll <inputFile> input.gal -compare input.gal . out.eval
+
+   
 ### -eval {: #eval }
 Evaluate a similarity matrix
@@ -35,7 +35,7 @@ A hacker&#39;s guide to building, editing, and running OpenBR.
         $ git clone https://github.com/biometrics/openbr.git
         $ cd openbr
-        $ git checkout 1.0
+        $ git checkout v1.1.0
         $ git submodule init
         $ git submodule update
@@ -116,7 +116,7 @@ A hacker&#39;s guide to building, editing, and running OpenBR.
         $ git clone https://github.com/biometrics/openbr.git
         $ cd openbr
-        $ git checkout 1.0
+        $ git checkout v1.1.0
         $ git submodule init
         $ git submodule update
@@ -196,7 +196,7 @@ A hacker&#39;s guide to building, editing, and running OpenBR.
             $ cd /c
             $ git clone https://github.com/biometrics/openbr.git
             $ cd openbr
-            $ git checkout 1.0
+            $ git checkout v1.1.0
             $ git submodule init
             $ git submodule update
@@ -277,7 +277,7 @@ A hacker&#39;s guide to building, editing, and running OpenBR.
         $ git clone https://github.com/biometrics/openbr.git
         $ cd openbr
-        $ git checkout 1.0
+        $ git checkout v1.1.0
         $ git submodule init
         $ git submodule update
@@ -131,23 +131,16 @@ static CvMat* cvPreprocessIndexArray( const CvMat* idx_arr, int data_arr_size, b
 //------------------------------------- FeatureEvaluator ---------------------------------------
-void FeatureEvaluator::init(Representation *_representation, int _maxSampleCount, int channels)
+void FeatureEvaluator::init(Representation *_representation, int _maxSampleCount)
 {
     representation = _representation;
-
-    int dx, dy;
-    Size windowSize = representation->windowSize(&dx, &dy);
-    data.create((int)_maxSampleCount, (windowSize.width + dx) * (windowSize.height + dy), CV_32SC(channels));
     cls.create( (int)_maxSampleCount, 1, CV_32FC1 );
 }
-void FeatureEvaluator::setImage(const Mat &img, uchar clsLabel, int idx)
+void FeatureEvaluator::setImage(const Template &src, uchar clsLabel, int idx)
 {
     cls.ptr<float>(idx)[0] = clsLabel;
-
-    Mat pp;
-    representation->preprocess(img, pp);
-    pp.reshape(0, 1).copyTo(data.row(idx));
+    data.append(representation->preprocess(src));
 }
 //----------------------------- CascadeBoostParams -------------------------------------------------
@@ -2,7 +2,6 @@
 #define _BOOST_H_
 #include "ml.h"
-#include <time.h>
 #include <openbr/openbr_plugin.h>
 #ifdef _WIN32
@@ -17,9 +16,9 @@ namespace br
 struct FeatureEvaluator
 {
     ~FeatureEvaluator() {}
-    void init(Representation *_representation, int _maxSampleCount, int channels);
-    void setImage(const cv::Mat& img, uchar clsLabel, int idx);
-    float operator()(int featureIdx, int sampleIdx) const { return representation->evaluate(data.row(sampleIdx), featureIdx); }
+    void init(Representation *_representation, int _maxSampleCount);
+    void setImage(const Template &src, uchar clsLabel, int idx);
+    float operator()(int featureIdx, int sampleIdx) const { return representation->evaluate(data[sampleIdx], featureIdx); }
     int getNumFeatures() const { return representation->numFeatures(); }
     int getMaxCatCount() const { return representation->maxCatCount(); }
@@ -27,7 +26,8 @@ struct FeatureEvaluator
     const cv::Mat& getCls() const { return cls; }
     float getCls(int si) const { return cls.at<float>(si, 0); }
-    cv::Mat data, cls;
+    cv::Mat cls;
+    TemplateList data;
     Representation *representation;
 };
@@ -825,16 +825,25 @@ static QStringList computeDetectionResults(const QList&lt;ResolvedDetection&gt; &amp;detec
                 if (prevFP / numImages < 0.1 && FP / numImages > 0.1 && discrete) {
                     qDebug("TAR @ FAR => %f : 0.1", TP / totalTrueDetections);
                     qDebug("Confidence: %f", detection.confidence);
+                    qDebug("TP vs. FP: %f to %f", TP, FP);
                 } else if (prevFP / numImages < 0.01 && FP / numImages > 0.01 && discrete) {
                     qDebug("TAR @ FAR => %f : 0.01", TP / totalTrueDetections);
                     qDebug("Confidence: %f", detection.confidence);
+                    qDebug("TP vs. FP: %f to %f", TP, FP);
+                } else if (prevFP / numImages < 0.001 && FP / numImages > 0.001 && discrete) {
+                    qDebug("TAR @ FAR => %f : 0.001", TP / totalTrueDetections);
+                    qDebug("Confidence: %f", detection.confidence);
+                    qDebug("TP vs. FP: %f to %f", TP, FP);
                 }
+
                 points.append(DetectionOperatingPoint(TP, FP, totalTrueDetections, numImages));
                 prevFP = FP;
             }
         }
     }
+    if (discrete) qDebug("Total TP vs. FP: %f to %f", TP, FP);
+
     const int keep = qMin(points.size(), Max_Points);
     if (keep < 1) qFatal("Insufficient points.");
@@ -1336,7 +1345,7 @@ void readKNNTruth(size_t probeCount, QVector&lt; QList&lt;size_t&gt; &gt; &amp;groundTruth, cons
         qFatal("Invalid ground truth file!");
 }
-void EvalKNN(const QString &knnGraph, const QString &knnTruth, const QString &iet)
+void EvalKNN(const QString &knnGraph, const QString &knnTruth, const QString &csv)
 {
     qDebug("Evaluating k-NN of %s against %s", qPrintable(knnGraph), qPrintable(knnTruth));
@@ -1406,19 +1415,14 @@ void EvalKNN(const QString &amp;knnGraph, const QString &amp;knnTruth, const QString &amp;ie
     if (numUnmatedSearches == 0)
         qFatal("No unmated searches!");
-    printf("Rank-%i Return Rate: %g\n", 1, getCMC(firstGenuineReturns, 1, numMatedSearches));
+
+    qDebug("Rank-%d Return Rate: %.3f", 1, getCMC(firstGenuineReturns, 1, numMatedSearches));
     if (k >=5)
-        printf("Rank-%i Return Rate: %g\n", 5, getCMC(firstGenuineReturns, 5, numMatedSearches));
+        qDebug("Rank-%d Return Rate: %.3f", 5, getCMC(firstGenuineReturns, 5, numMatedSearches));
     if (k >=10)
-        printf("Rank-%i Return Rate: %g\n", 10, getCMC(firstGenuineReturns, 10, numMatedSearches));
-
-    printf("Rank-%zu Return Rate: %g\n", k, double(numMatedSimilarities) / double(numMatedSearches));
+        qDebug("Rank-%d Return Rate: %.3f", 10, getCMC(firstGenuineReturns, 10, numMatedSearches));
-    // Open the output file
-    QFile ietFile(iet);
-    if (!ietFile.open(QFile::WriteOnly | QFile::Text))
-        qFatal("Failed to open IET file for writing!");
-    ietFile.write("Threshold,FPIR,FNIR\n");
+    qDebug("Rank-%zu Return Rate: %.3f", k, double(numMatedSimilarities) / double(numMatedSearches));
     /*
      * Iterate through the similarity scores highest-to-lowest,
@@ -1447,10 +1451,28 @@ void EvalKNN(const QString &amp;knnGraph, const QString &amp;knnTruth, const QString &amp;ie
         }
     }
-    foreach(const OperatingPoint &operatingPoint, operatingPoints)
-        ietFile.write(qPrintable(QString::number(operatingPoint.score) + "," +
-                                 QString::number(operatingPoint.FAR) + "," +
-                                 QString::number(operatingPoint.TAR) + "\n"));
+    if (!csv.isEmpty()) {
+        // Open the output file
+        QFile ietFile(csv);
+        if (!ietFile.open(QFile::WriteOnly | QFile::Text))
+            qFatal("Failed to open IET file for writing!");
+        ietFile.write("Plot,X,Y,Z\n");
+        // Write CMC
+        const int Max_Retrieval = min(200, (int)k);
+        for (int i=1; i<=Max_Retrieval; i++) {
+            const float retrievalRate = getCMC(firstGenuineReturns, i, numMatedSearches);
+            ietFile.write(qPrintable(QString("CMC,%1,%2,0\n").arg(QString::number(i), QString::number(retrievalRate))));
+        }
+
+        foreach(const OperatingPoint &operatingPoint, operatingPoints)
+            ietFile.write(qPrintable("IET," +
+                                     QString::number(operatingPoint.FAR) + "," +
+                                     QString::number(operatingPoint.TAR) + "," +
+                                     QString::number(operatingPoint.score) + "\n"));
+    }
+
+    qDebug("FNIR @ FPIR = 0.1:   %.3f", 1-getOperatingPointGivenFAR(operatingPoints, 0.1).TAR);
+    qDebug("FNIR @ FPIR = 0.01:  %.3f", 1-getOperatingPointGivenFAR(operatingPoints, 0.01).TAR);
 }
 } // namespace br
@@ -33,7 +33,7 @@ namespace br
     float EvalDetection(const QString &predictedGallery, const QString &truthGallery, const QString &csv = "", bool normalize = false, int minSize = 0, int maxSize = 0); // Return average overlap
     float EvalLandmarking(const QString &predictedGallery, const QString &truthGallery, const QString &csv = "", int normalizationIndexA = 0, int normalizationIndexB = 1, int sampleIndex = 0, int totalExamples = 5); // Return average error
     void EvalRegression(const QString &predictedGallery, const QString &truthGallery, QString predictedProperty = "", QString truthProperty = "");
-    void EvalKNN(const QString &knnGraph, const QString &knnTruth, const QString &iet);
+    void EvalKNN(const QString &knnGraph, const QString &knnTruth, const QString &csv = "");
     struct Candidate
     {
@@ -436,7 +436,7 @@ public:
 };
 // TODO: Make sure case where no confidences are inputted works.
-void OpenCVUtils::group(QList<Rect> &rects, QList<float> &confidences, float confidenceThreshold, float epsilon)
+void OpenCVUtils::group(QList<Rect> &rects, QList<float> &confidences, float confidenceThreshold, int minNeighbors, float epsilon)
 {
     if (rects.isEmpty())
         return;
@@ -450,7 +450,7 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
     vector<Rect> rrects(nClasses);
     // Total number of rects in each class
-    vector<int> rweights(nClasses, 0);
+    vector<int> neighbors(nClasses, 0);
     vector<float> rejectWeights(nClasses, -std::numeric_limits<float>::max());
     for (size_t i = 0; i < labels.size(); i++)
@@ -460,7 +460,7 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
         rrects[cls].y += rects[i].y;
         rrects[cls].width += rects[i].width;
         rrects[cls].height += rects[i].height;
-        rweights[cls]++;
+        neighbors[cls]++;
     }
     if (useConfidences)
@@ -478,7 +478,7 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
     for (int i = 0; i < nClasses; i++)
     {
         Rect r = rrects[i];
-        float s = 1.f/rweights[i];
+        float s = 1.f/neighbors[i];
         rrects[i] = Rect(saturate_cast<int>(r.x*s),
              saturate_cast<int>(r.y*s),
              saturate_cast<int>(r.width*s),
@@ -488,7 +488,7 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
     rects.clear();
     confidences.clear();
-    // Aggregate by comparing average rectangles against other average rectangels
+    // Aggregate by comparing average rectangles against other average rectangles
     for (int i = 0; i < nClasses; i++)
     {
         // Average rectangle
@@ -496,19 +496,22 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
         // Used to eliminate rectangles with few neighbors in the case of no weights
         // int n1 = levelWeights ? rejectLevels[i] : rweights[i];
-        float w1 = rejectWeights[i];
+        const float w1 = rejectWeights[i];
         // Eliminate rectangle if it doesn't meet confidence criteria
         if (w1 <= confidenceThreshold)
             continue;
+        const int n1 = neighbors[i];
+        if (n1 < minNeighbors)
+            continue;
+
         // filter out small face rectangles inside large rectangles
         int j;
         for (j = 0; j < nClasses; j++)
         {
-            float w2 = rejectWeights[j];
-
-            if (j == i)
+            const int n2 = neighbors[j];
+            if (j == i || n2 < minNeighbors)
                 continue;
             Rect r2 = rrects[j];
@@ -516,13 +519,11 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
             int dx = saturate_cast<int>(r2.width * epsilon);
             int dy = saturate_cast<int>(r2.height * epsilon);
+            float w2 = rejectWeights[j];
+
             // If, r1 is within the r2 AND
-            // the second rectangle reaches a later stage than the first
-            // where both the first and the second must have a stage greater than three OR
-            // the first doens't reach the third stage.
-            // Changeto: second rectangle has a higher confidence than the first OR
-            // the first has a low confidence.
-            // Then, eliminate the first rectangle.
+            // r2 has a higher confidence than r1
+            // then, eliminate the r1
             if(r1.x >= r2.x - dx &&
                r1.y >= r2.y - dy &&
                r1.x + r1.width <= r2.x + r2.width + dx &&
@@ -531,7 +532,6 @@ void OpenCVUtils::group(QList&lt;Rect&gt; &amp;rects, QList&lt;float&gt; &amp;confidences, float con
                break;
         }
-        // Need to return rects and confidences
         if( j == nClasses )
         {
             rects.append(r1);
@@ -102,7 +102,7 @@ namespace OpenCVUtils
     float overlap(const QRectF &rect1, const QRectF &rect2);
     // Misc
-    void group(QList<cv::Rect> &rects, QList<float> &confidences, float confidenceThreshold, float epsilon);
+    void group(QList<cv::Rect> &rects, QList<float> &confidences, float confidenceThreshold, int minNeighbors, float epsilon);
     void flip(const br::Template &src, br::Template &dst, int axis, bool flipMat=true, bool flipPoints=true, bool flipRects=true);
     void flip(const br::TemplateList &src, br::TemplateList &dst, int axis, bool flipMat=true, bool flipPoints=true, bool flipRects=true);
@@ -346,4 +346,32 @@ bool PlotMetadata(const QStringList &amp;files, const QString &amp;columns, bool show)
     return p.finalize(show);
 }
+bool PlotKNN(const QStringList &files, const File &destination, bool show)
+{
+    qDebug("Plotting %d k-NN file(s) to %s", files.size(), qPrintable(destination));
+    RPlot p(files, destination);
+    p.file.write("\nformatData(type=\"knn\")\n\n");
+
+    QMap<QString,File> optMap;
+    optMap.insert("rocOptions", File(QString("[xTitle=False Positive Identification Rate (FPIR),yTitle=True Positive Identification Rate (TPIR),xLog=true,yLog=false]")));
+    optMap.insert("ietOptions", File(QString("[xTitle=False Positive Identification Rate (FPIR),yTitle=False Negative Identification Rate (FNIR),xLog=true,yLog=true]")));
+    optMap.insert("cmcOptions", File(QString("[xTitle=Rank,yTitle=Retrieval Rate,xLog=true,yLog=false,size=1,xLabels=(1,5,10,50,100),xBreaks=(1,5,10,50,100)]")));
+
+    foreach (const QString &key, optMap.keys()) {
+        const QStringList options = destination.get<QStringList>(key, QStringList());
+        foreach (const QString &option, options) {
+            QStringList words = QtUtils::parse(option, '=');
+            QtUtils::checkArgsSize(words[0], words, 1, 2);
+            optMap[key].set(words[0], words[1]);
+        }
+    }
+
+    QString plot = "plot <- plotLine(lineData=%1, options=list(%2), flipY=%3)\nplot\n";
+    p.file.write(qPrintable(QString(plot).arg("IET", toRList(optMap["rocOptions"]), "TRUE")));
+    p.file.write(qPrintable(QString(plot).arg("IET", toRList(optMap["ietOptions"]), "FALSE")));
+    p.file.write(qPrintable(QString(plot).arg("CMC", toRList(optMap["cmcOptions"]), "FALSE")));
+
+    return p.finalize(show);
+}
+
 } // namespace br
@@ -28,6 +28,7 @@ namespace br
     bool PlotDetection(const QStringList &files, const File &destination, bool show = false);
     bool PlotLandmarking(const QStringList &files, const File &destination, bool show = false);
     bool PlotMetadata(const QStringList &files, const QString &destination, bool show = false);
+    bool PlotKNN(const QStringList &files, const File &destination, bool show = false);
 }
 #endif // BR_PLOT_H
@@ -145,9 +145,9 @@ void br_eval_regression(const char *predicted_gallery, const char *truth_gallery
     EvalRegression(predicted_gallery, truth_gallery, predicted_property, truth_property);
 }
-void br_eval_knn(const char *knnGraph, const char *knnTruth, const char *iet)
+void br_eval_knn(const char *knnGraph, const char *knnTruth, const char *csv)
 {
-    EvalKNN(knnGraph, knnTruth, iet);
+    EvalKNN(knnGraph, knnTruth, csv);
 }
 void br_finalize()
@@ -221,6 +221,11 @@ bool br_plot_metadata(int num_files, const char *files[], const char *columns, b
     return PlotMetadata(QtUtils::toStringList(num_files, files), columns, show);
 }
+bool br_plot_knn(int num_files, const char *files[], const char *destination, bool show)
+{
+    return PlotKNN(QtUtils::toStringList(num_files, files), destination, show);
+}
+
 float br_progress()
 {
     return Globals->progress();
@@ -64,7 +64,7 @@ BR_EXPORT float br_eval_landmarking(const char *predicted_gallery, const char *t
 BR_EXPORT void br_eval_regression(const char *predicted_gallery, const char *truth_gallery, const char *predicted_property = "", const char *truth_property = "");
-BR_EXPORT void br_eval_knn(const char *knnGraph, const char *knnTruth, const char *iet);
+BR_EXPORT void br_eval_knn(const char *knnGraph, const char *knnTruth, const char *csv = "");
 BR_EXPORT void br_finalize();
@@ -93,6 +93,8 @@ BR_EXPORT bool br_plot_landmarking(int num_files, const char *files[], const cha
 BR_EXPORT bool br_plot_metadata(int num_files, const char *files[], const char *columns, bool show = false);
+BR_EXPORT bool br_plot_knn(int num_files, const char *files[], const char *destination, bool show = false);
+
 BR_EXPORT float br_progress();
 BR_EXPORT void br_read_pipe(const char *pipe, int *argc, char ***argv);
@@ -882,13 +882,11 @@ public:
     static Representation *make(QString str, QObject *parent); /*!< \brief Make a representation from a string. */
-    virtual void preprocess(const cv::Mat &src, cv::Mat &dst) const { dst = src; }
-    virtual void train(const QList<cv::Mat> &images, const QList<float> &labels) { (void) images; (void)labels; }
-
-    virtual float evaluate(const cv::Mat &image, int idx) const = 0;
-    // By convention, an empty indices list will result in all feature responses being calculated
-    // and returned.
-    virtual cv::Mat evaluate(const cv::Mat &image, const QList<int> &indices = QList<int>()) const = 0;
+    virtual Template preprocess(const Template &src) const { return src; }
+    virtual void train(const TemplateList &data) { (void)data; }
+    virtual float evaluate(const Template &src, int idx) const = 0;
+    // By convention passing an empty list evaluates all features in the representation
+    virtual cv::Mat evaluate(const Template &src, const QList<int> &indices = QList<int>()) const = 0;
     virtual cv::Size windowSize(int *dx = NULL, int *dy = NULL) const = 0; // dx and dy should indicate the change to the original window size after preprocessing
     virtual int numChannels() const { return 1; }
@@ -905,13 +903,13 @@ public:
     static Classifier *make(QString str, QObject *parent);
-    virtual void train(const QList<cv::Mat> &images, const QList<float> &labels) = 0;
-    virtual float classify(const cv::Mat &image, bool process = true, float *confidence = NULL) const = 0;
+    virtual void train(const TemplateList &data) { (void)data; }
+    virtual float classify(const Template &src, bool process = true, float *confidence = NULL) const = 0;
     // Slots for representations
-    virtual cv::Mat preprocess(const cv::Mat &image) const = 0;
+    virtual Template preprocess(const Template &src) const { return src; }
     virtual cv::Size windowSize(int *dx = NULL, int *dy = NULL) const = 0;
-    virtual int numFeatures() const = 0;
+    virtual int numFeatures() const { return 0; }
 };
@@ -125,20 +125,20 @@ private:
     QList<Node*> classifiers;
     float threshold;
-    void train(const QList<Mat> &images, const QList<float> &labels)
+    void train(const TemplateList &data)
     {
-        representation->train(images, labels);
+        representation->train(data);
         CascadeBoostParams params(type, minTAR, maxFAR, trimRate, maxDepth, maxWeakCount);
         FeatureEvaluator featureEvaluator;
-        featureEvaluator.init(representation, images.size(), representation->numChannels());
+        featureEvaluator.init(representation, data.size());
-        for (int i = 0; i < images.size(); i++)
-            featureEvaluator.setImage(images[i], labels[i], i);
+        for (int i = 0; i < data.size(); i++)
+            featureEvaluator.setImage(data[i], data[i].file.get<float>("Label"), i);
         CascadeBoost boost;
-        boost.train(&featureEvaluator, images.size(), 1024, 1024, representation->numChannels(), params);
+        boost.train(&featureEvaluator, data.size(), 1024, 1024, representation->numChannels(), params);
         threshold = boost.getThreshold();
@@ -149,13 +149,9 @@ private:
         }
     }
-    float classify(const Mat &image, bool process, float *confidence) const
+    float classify(const Template &src, bool process, float *confidence) const
     {
-        Mat m;
-        if (process)
-            m = preprocess(image);
-        else
-            m = image;
+        Template t = process ? preprocess(src) : src;
         float sum = 0;
         for (int i = 0; i < classifiers.size(); i++) {
@@ -163,10 +159,10 @@ private:
             while (node->left) {
                 if (representation->maxCatCount() > 0) {
-                    int c = (int)representation->evaluate(m, node->featureIdx);
+                    int c = (int)representation->evaluate(t, node->featureIdx);
                     node = (node->subset[c >> 5] & (1 << (c & 31))) ? node->left : node->right;
                 } else {
-                    double val = representation->evaluate(m, node->featureIdx);
+                    double val = representation->evaluate(t, node->featureIdx);
                     node = val <= node->threshold ? node->left : node->right;
                 }
             }
@@ -184,11 +180,9 @@ private:
         return representation->numFeatures();
     }
-    Mat preprocess(const Mat &image) const
+    Template preprocess(const Template &src) const
     {
-        Mat dst;
-        representation->preprocess(image, dst);
-        return dst;
+        return representation->preprocess(src);
     }
     Size windowSize(int *dx, int *dy) const
 #include <opencv2/imgproc/imgproc.hpp>
-#include <opencv2/highgui/highgui.hpp>
+
 #include <openbr/plugins/openbr_internal.h>
 #include <openbr/core/common.h>
+#include "openbr/core/opencvutils.h"
 #include <QtConcurrent>
@@ -37,7 +38,7 @@ struct Miner
     Mat mine(bool *newImg)
     {
         // Copy region of winSize region of img into m
-        Mat window(windowSize.height, windowSize.width, CV_8UC1,
+        Mat window(windowSize.height, windowSize.width, CV_8U,
                    (void*)(scaledSrc.data + point.y * scaledSrc.step + point.x * scaledSrc.elemSize()),
                    scaledSrc.step);
@@ -100,31 +101,30 @@ class CascadeClassifier : public Classifier
     BR_PROPERTY(bool, requireAllStages, false)
     QList<Classifier *> stages;
-    QList<Mat> posImages, negImages;
-    QList<Mat> posSamples, negSamples;
+    TemplateList posImages, negImages;
+    TemplateList posSamples, negSamples;
     QList<int> indices;
     int negIndex, posIndex, samplingRound;
-    QMutex samplingMutex, miningMutex, passedMutex;
+    QMutex samplingMutex, miningMutex;
     void init()
     {
         negIndex = posIndex = samplingRound = 0;
     }
-    bool getPositive(Mat &img)
+    bool getPositive(Template &img)
     {
         if (posIndex >= posImages.size())
             return false;
-
-        posImages[indices[posIndex++]].copyTo(img);
+        img = posImages[indices[posIndex++]];
         return true;
     }
-    Mat getNegative(Point &offset)
+    Template getNegative(Point &offset)
     {
-        Mat negative;
+        Template negative;
         const Size size = windowSize();
         // Grab negative from list
@@ -136,9 +136,9 @@ class CascadeClassifier : public Classifier
             samplingRound = samplingRound % (size.width * size.height);
             negIndex %= count;
-            offset.x = qMin( (int)samplingRound % size.width, negative.cols - size.width);
-            offset.y = qMin( (int)samplingRound / size.width, negative.rows - size.height);
-            if (!negative.empty() && negative.type() == CV_8UC1
+            offset.x = qMin( (int)samplingRound % size.width, negative.m().cols - size.width);
+            offset.y = qMin( (int)samplingRound / size.width, negative.m().rows - size.height);
+            if (!negative.m().empty() && negative.m().type() == CV_8U
                     && offset.x >= 0 && offset.y >= 0)
                 break;
         }
@@ -146,35 +146,32 @@ class CascadeClassifier : public Classifier
         return negative;
     }
-    int mine()
+    uint64 mine()
     {
-        int passedNegatives = 0;
+        uint64 passedNegatives = 0;
         forever {
-            Mat negative;
+            Template negative;
             Point offset;
-            samplingMutex.lock();
+            QMutexLocker samplingLocker(&samplingMutex);
             negative = getNegative(offset);
-            samplingMutex.unlock();
+            samplingLocker.unlock();
-            Miner miner(negative, windowSize(), offset);
+            Miner miner(negative.m(), windowSize(), offset);
             forever {
                 bool newImg;
-                Mat sample = miner.mine(&newImg);
+                Template sample(negative.file, miner.mine(&newImg));
                 if (!newImg) {
                     if (negSamples.size() >= numNegs)
                         return passedNegatives;
                     float confidence;
                     if (classify(sample, true, &confidence) != 0) {
-                        miningMutex.lock();
-                        if (negSamples.size() >= numNegs) {
-                            miningMutex.unlock();
+                        QMutexLocker miningLocker(&miningMutex);
+                        if (negSamples.size() >= numNegs)
                             return passedNegatives;
-                        }
                         negSamples.append(sample);
                         printf("Negative samples: %d\r", negSamples.size());
-                        miningMutex.unlock();
                     }
                     passedNegatives++;
@@ -184,16 +181,16 @@ class CascadeClassifier : public Classifier
         }
     }
-    void train(const QList<Mat> &images, const QList<float> &labels)
+    void train(const TemplateList &data)
     {
-        for (int i = 0; i < images.size(); i++)
-            labels[i] == 1 ? posImages.append(images[i]) : negImages.append(images[i]);
+        foreach (const Template &t, data)
+            t.file.get<float>("Label") == 1.0f ? posImages.append(t) : negImages.append(t);
-        qDebug() << "Total images:" << images.size()
+        qDebug() << "Total images:" << data.size()
                  << "\nTotal positive images:" << posImages.size()
                  << "\nTotal negative images:" << negImages.size();
-        indices = Common::RandSample(posImages.size(),posImages.size(),true);
+        indices = Common::RandSample(posImages.size(), posImages.size(), true);
         stages.reserve(numStages);
         for (int i = 0; i < numStages; i++) {
@@ -212,27 +209,17 @@ class CascadeClassifier : public Classifier
                 return;
             }
-            QList<float> posLabels;
-            posLabels.reserve(posSamples.size());
-            for (int j=0; j<posSamples.size(); j++)
-                posLabels.append(1);
-
-            QList<float> negLabels;
-            negLabels.reserve(negSamples.size());
-            for (int j=0; j<negSamples.size(); j++)
-                negLabels.append(0);
-
-            stages[i]->train(posSamples+negSamples, posLabels+negLabels);
+            stages[i]->train(posSamples + negSamples);
             qDebug() << "END>";
         }
     }
-    float classify(const Mat &image, bool process, float *confidence) const
+    float classify(const Template &src, bool process, float *confidence) const
     {
         float stageConf = 0.0f;
         foreach (const Classifier *stage, stages) {
-            float result = stage->classify(image, process, &stageConf);
+            float result = stage->classify(src, process, &stageConf);
             if (confidence)
                 *confidence += stageConf;
             if (result == 0.0f)
@@ -246,9 +233,9 @@ class CascadeClassifier : public Classifier
         return stages.first()->numFeatures();
     }
-    Mat preprocess(const Mat &image) const
+    Template preprocess(const Template &src) const
     {
-        return stages.first()->preprocess(image);
+        return stages.first()->preprocess(src);
     }
     Size windowSize(int *dx = NULL, int *dy = NULL) const
@@ -276,16 +263,13 @@ class CascadeClassifier : public Classifier
 private:
     float getSamples()
     {
-        posSamples.clear();
-        posSamples.reserve(numPos);
-        negSamples.clear();
-        negSamples.reserve(numNegs);
+        posSamples.clear(); posSamples.reserve(numPos);
+        negSamples.clear(); negSamples.reserve(numNegs);
         posIndex = 0;
         float confidence;
         while (posSamples.size() < numPos) {
-            Mat pos(windowSize(), CV_8UC1);
-
+            Template pos;
             if (!getPositive(pos))
                 qFatal("Cannot get another positive sample!");
@@ -297,13 +281,13 @@ private:
         qDebug() << "POS count : consumed  " << posSamples.size() << ":" << posIndex;
-        QFutureSynchronizer<int> futures;
+        QFutureSynchronizer<uint64> futures;
         for (int i=0; i<QThread::idealThreadCount(); i++)
             futures.addFuture(QtConcurrent::run(this, &CascadeClassifier::mine));
         futures.waitForFinished();
-        int passedNegs = 0;
-        QList<QFuture<int> > results = futures.futures();
+        uint64 passedNegs = 0;
+        QList<QFuture<uint64> > results = futures.futures();
         for (int i=0; i<results.size(); i++)
             passedNegs += results[i].result();
+#include <opencv2/imgproc/imgproc.hpp>
 #include <dlib/image_processing/frontal_face_detector.h>
 #include <dlib/opencv.h>
@@ -45,14 +46,18 @@ private:
     {
         dst = src;
-        shape_predictor *sp = shapeResource.acquire();
+        shape_predictor *const sp = shapeResource.acquire();
-        cv_image<unsigned char> cimg(src.m());
+        cv::Mat cvImage = src.m();
+        if (cvImage.channels() == 3)
+            cv::cvtColor(cvImage, cvImage, CV_BGR2GRAY);
+
+        cv_image<unsigned char> cimg(cvImage);
         array2d<unsigned char> image;
         assign_image(image,cimg);
         if (src.file.rects().isEmpty()) { //If the image has no rects assume the whole image is a face
-            rectangle r(0, 0, src.m().cols, src.m().rows);
+            rectangle r(0, 0, cvImage.cols, cvImage.rows);
             full_object_detection shape = (*sp)(image, r);
             for (size_t i = 0; i < shape.num_parts(); i++)
                 dst.file.appendPoint(QPointF(shape.part(i)(0),shape.part(i)(1)));
@@ -138,7 +138,7 @@ private:
             param.weight = NULL;
         }
-        m = *train_svm(&prob, &param);
+        //m = *train_svm(&prob, &param);
         delete[] param.weight;
         delete[] param.weight_label;
@@ -39,7 +39,7 @@ class AlgorithmsInitializer : public Initializer
         Globals->abbreviations.insert("FR_Mouth", "(CropFromLandmarks([59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76])+Resize(24,48))");
         Globals->abbreviations.insert("FR_Nose", "(CropFromLandmarks([16,17,18,19,20,21,22,23,24,25,26,27],padding=3)+Resize(36,36))");
         Globals->abbreviations.insert("FR_Face", "(Crop(24,24,88,88)+Resize(44,44))");
-        Globals->abbreviations.insert("FR_Detect", "(FaceDetection+Stasm+Rename(StasmLeftEye,Affine_1,true)+Rename(StasmRightEye,Affine_0,true)+Affine(136,136,0.35,0.35,warpPoints=true))");
+        Globals->abbreviations.insert("FR_Detect", "(Open+Cvt(Gray)+Cascade+Stasm+Rename(StasmLeftEye,Affine_1,true)+Rename(StasmRightEye,Affine_0,true)+Affine(136,136,0.35,0.35,warpPoints=true))");
         Globals->abbreviations.insert("FR_Represent", "((DenseHOG/DenseLBP)+Cat+LDA(.98)+Normalize(L2))");
         Globals->abbreviations.insert("GenderClassification", "FaceDetection+Expand+FaceClassificationRegistration+Expand+<FaceClassificationExtraction>+<GenderClassifier>+Discard");
@@ -76,12 +76,18 @@ class UnitDistance : public Distance
     float compare(const Template &target, const Template &query) const
     {
-        return a * (distance->compare(target, query) - b);
+        return normalize(distance->compare(target, query));
     }
     float compare(const cv::Mat &target, const cv::Mat &query) const
     {
-        return a * (distance->compare(target, query) - b);
+        return normalize(distance->compare(target, query));
+    }
+	
+	float normalize(float score) const
+    {
+        if (!Globals->scoreNormalization) return score;
+        return a * (score - b);
     }
 };
@@ -26,6 +26,12 @@ extern &quot;C&quot;
 #include <libswscale/swscale.h>
 }
+// http://stackoverflow.com/questions/24057248/ffmpeg-undefined-references-to-av-frame-alloc
+#if LIBAVCODEC_VERSION_INT < AV_VERSION_INT(55,28,1)
+#define av_frame_alloc avcodec_alloc_frame
+#define av_frame_free avcodec_free_frame
+#endif
+
 using namespace cv;
 namespace br
@@ -18,6 +18,8 @@
 #include <openbr/plugins/openbr_internal.h>
+using namespace cv;
+
 namespace br
 {
@@ -32,7 +34,20 @@ class EqualizeHistTransform : public UntrainableTransform
     void project(const Template &src, Template &dst) const
     {
-        cv::equalizeHist(src, dst);
+        if (src.m().channels() == 1) {
+            equalizeHist(src, dst);
+        } else if (src.m().channels() == 3) {
+            // http://stackoverflow.com/questions/15007304/histogram-equalization-not-working-on-color-image-opencv
+            Mat ycrcb;
+            cvtColor(src, ycrcb, CV_BGR2YCrCb);
+            vector<Mat> channels;
+            split(ycrcb, channels);
+            equalizeHist(channels[0], channels[0]);
+            merge(channels, ycrcb);
+            cvtColor(ycrcb, dst, CV_YCrCb2BGR);
+        } else {
+            qFatal("Invalid channel count!");
+        }
     }
 };
+#include <opencv2/imgproc/imgproc.hpp>
 #include <openbr/plugins/openbr_internal.h>
 using namespace cv;
@@ -5,24 +6,36 @@ using namespace cv;
 namespace br
 {
+/*!
+ * \ingroup transforms
+ * \brief Pads an image.
+ * \author Scott Klum \cite sklum
+ */
 class PadTransform : public UntrainableTransform
 {
     Q_OBJECT
-
-    Q_PROPERTY(int padSize READ get_padSize WRITE set_padSize RESET reset_padSize STORED false)
-    Q_PROPERTY(int padValue READ get_padValue WRITE set_padValue RESET reset_padValue STORED false)
-    BR_PROPERTY(int, padSize, 0)
-    BR_PROPERTY(int, padValue, 0)
+    Q_ENUMS(Method)
+
+public:
+    /*!< */
+    enum Border { Replicate = BORDER_REPLICATE,
+                  Reflect = BORDER_REFLECT_101,
+                  Constant = BORDER_CONSTANT};
+
+private:
+    Q_PROPERTY(Border border READ get_border WRITE set_border RESET reset_border STORED false)
+    Q_PROPERTY(float percent READ get_percent WRITE set_percent RESET reset_percent STORED false)
+    Q_PROPERTY(int value READ get_value WRITE set_value RESET reset_value STORED false)
+    BR_PROPERTY(Border, border, Replicate)
+    BR_PROPERTY(float, percent, .1)
+    BR_PROPERTY(float, value, 0)
     void project(const Template &src, Template &dst) const
     {
-        dst.file = src.file;
-
-        foreach (const Mat &m, src) {
-            Mat padded = padValue * Mat::ones(m.rows + 2*padSize, m.cols + 2*padSize, m.type());
-            padded(Rect(padSize, padSize, padded.cols - padSize, padded.rows - padSize)) = m;
-            dst += padded;
-        }
+        int top, bottom, left, right;
+        top = percent*src.m().rows; bottom = percent*src.m().rows;
+        left = percent*src.m().cols; right = percent*src.m().cols;
+        copyMakeBorder(src, dst, top, bottom, left, right, border, Scalar(value));
     }
 };
@@ -55,9 +55,15 @@ private:
     void project(const Template &src, Template &dst) const
     {
-        if (!preserveAspect)
+        if (!preserveAspect) {
             resize(src, dst, Size((columns == -1) ? src.m().cols*rows/src.m().rows : columns, rows), 0, 0, method);
-        else {
+            const float rowScaleFactor = (float)rows/src.m().rows;
+            const float colScaleFactor = (float)columns/src.m().cols;
+            QList<QPointF> points = src.file.points();
+            for (int i=0; i<points.size(); i++)
+                points[i] = QPointF(points[i].x() * colScaleFactor,points[i].y() * rowScaleFactor);
+            dst.file.setPoints(points);
+        } else {
             float inRatio = (float) src.m().rows / src.m().cols;
             float outRatio = (float) rows / columns;
             dst = Mat::zeros(rows, columns, src.m().type());
@@ -16,6 +16,7 @@
 #include <opencv2/imgproc/imgproc.hpp>
 #include <openbr/plugins/openbr_internal.h>
+#include <openbr/core/opencvutils.h>
 using namespace cv;
@@ -32,12 +33,14 @@ class RndRotateTransform : public UntrainableTransform
     Q_OBJECT
     Q_PROPERTY(QList<int> range READ get_range WRITE set_range RESET reset_range STORED false)
+    Q_PROPERTY(int center READ get_center WRITE set_center RESET reset_center STORED false)
     BR_PROPERTY(QList<int>, range, QList<int>() << -15 << 15)
+    BR_PROPERTY(int, center, -1)
     void project(const Template &src, Template &dst) const {
         int span = range.first() - range.last();
-        int angle = (rand() % span) + range.first();
-        Mat rotMatrix = getRotationMatrix2D(Point2f(src.m().rows/2,src.m().cols/2),angle,1.0);
+        int angle = span == 0 ? range.first() : (rand() % span) + range.first();
+        Mat rotMatrix = getRotationMatrix2D(center == -1 ? Point2f(src.m().rows/2,src.m().cols/2) : OpenCVUtils::toPoint(src.file.points()[center]),angle,1.0);
         warpAffine(src,dst,rotMatrix,Size(src.m().cols,src.m().rows),INTER_LINEAR,BORDER_REFLECT_101);
         QList<QPointF> points = src.file.points();
@@ -34,8 +34,10 @@ class ROITransform : public UntrainableTransform
     Q_OBJECT
     Q_PROPERTY(QString propName READ get_propName WRITE set_propName RESET reset_propName STORED false)
     Q_PROPERTY(bool copyOnCrop READ get_copyOnCrop WRITE set_copyOnCrop RESET reset_copyOnCrop STORED false)
+    Q_PROPERTY(int shiftPoints READ get_shiftPoints WRITE set_shiftPoints RESET reset_shiftPoints STORED false)
     BR_PROPERTY(QString, propName, "")
     BR_PROPERTY(bool, copyOnCrop, false)
+    BR_PROPERTY(int, shiftPoints, -1)
     void project(const Template &src, Template &dst) const
     {
@@ -55,6 +57,16 @@ class ROITransform : public UntrainableTransform
             if (Globals->verbose)
                 qWarning("No rects present in file.");
         }
+
+        if (shiftPoints != -1) {
+            // Shift the points to the rect with the index provided
+            QRectF rect = src.file.rects()[shiftPoints];
+            QList<QPointF> points = src.file.points();
+            for (int i=0; i<points.size(); i++)
+                points[i] -= rect.topLeft();
+            dst.file.setPoints(points);
+        }
+
         dst.file.clearRects();
         if (copyOnCrop)
@@ -17,7 +17,6 @@
 #include <openbr/plugins/openbr_internal.h>
 #include <openbr/core/opencvutils.h>
 #include <openbr/core/qtutils.h>
-#include <opencv2/highgui/highgui.hpp>
 #include <opencv2/imgproc/imgproc.hpp>
@@ -35,7 +34,6 @@ namespace br
  * \br_property int minSize The smallest sized object to detect in pixels
  * \br_property int maxSize The largest sized object to detect in pixels. A negative value will set maxSize == image size
  * \br_property float scaleFactor The factor to scale the image by during each resize.
- * \br_property int minNeighbors Parameter for non-maximum supression
  * \br_property float confidenceThreshold A threshold for positive detections. Positive detections returned by the classifier that have confidences below this threshold are considered negative detections.
  * \br_property float eps Parameter for non-maximum supression
  */
@@ -50,6 +48,7 @@ class SlidingWindowTransform : public MetaTransform
     Q_PROPERTY(float scaleFactor READ get_scaleFactor WRITE set_scaleFactor RESET reset_scaleFactor STORED false)
     Q_PROPERTY(float confidenceThreshold READ get_confidenceThreshold WRITE set_confidenceThreshold RESET reset_confidenceThreshold STORED false)
     Q_PROPERTY(float eps READ get_eps WRITE set_eps RESET reset_eps STORED false)
+    Q_PROPERTY(float minNeighbors READ get_minNeighbors WRITE set_minNeighbors RESET reset_minNeighbors STORED false)
     BR_PROPERTY(br::Classifier*, classifier, NULL)
     BR_PROPERTY(int, minSize, 20)
@@ -57,10 +56,11 @@ class SlidingWindowTransform : public MetaTransform
     BR_PROPERTY(float, scaleFactor, 1.2)
     BR_PROPERTY(float, confidenceThreshold, 10)
     BR_PROPERTY(float, eps, 0.2)
+    BR_PROPERTY(int, minNeighbors, 3)
     void train(const TemplateList &data)
     {
-        classifier->train(data.data(), File::get<float>(data, "Label", -1));
+        classifier->train(data);
     }
     void project(const Template &src, Template &dst) const
@@ -123,15 +123,18 @@ class SlidingWindowTransform : public MetaTransform
                     Mat scaledImage(scaledImageSize, CV_8U, imageBuffer.data);
                     resize(m, scaledImage, scaledImageSize, 0, 0, CV_INTER_LINEAR);
-                    Mat repImage = classifier->preprocess(scaledImage);
+
+                    Template repImage(t.file, scaledImage);
+                    repImage = classifier->preprocess(repImage);
                     int step = factor > 2. ? 1 : 2;
                     for (int y = 0; y < processingRectSize.height; y += step) {
                         for (int x = 0; x < processingRectSize.width; x += step) {
-                            Mat window = repImage(Rect(Point(x, y), Size(originalWindowSize.width + dx, originalWindowSize.height + dy))).clone();
+                            Mat window = repImage.m()(Rect(Point(x, y), Size(originalWindowSize.width + dx, originalWindowSize.height + dy))).clone();
+                            Template t(window);
                             float confidence = 0;
-                            int result = classifier->classify(window, false, &confidence);
+                            int result = classifier->classify(t, false, &confidence);
                             if (result == 1) {
                                 rects.append(Rect(cvRound(x*factor), cvRound(y*factor), windowSize.width, windowSize.height));
@@ -146,7 +149,7 @@ class SlidingWindowTransform : public MetaTransform
                     }
                 }
-                OpenCVUtils::group(rects, confidences, confidenceThreshold, eps);
+                OpenCVUtils::group(rects, confidences, confidenceThreshold, minNeighbors, eps);
                 if (!enrollAll && rects.empty()) {
                     rects.append(Rect(0, 0, m.cols, m.rows));
@@ -14,23 +14,14 @@ class AveragePointsTransform : public UntrainableMetadataTransform
     Q_PROPERTY(QList<int> indices READ get_indices WRITE set_indices RESET reset_indices STORED false)
     Q_PROPERTY(QString metaName READ get_metaName WRITE set_metaName RESET reset_metaName STORED true)
     Q_PROPERTY(bool append READ get_append WRITE set_append RESET reset_append STORED true)
-    Q_PROPERTY(int nLandmarks READ get_nLandmarks WRITE set_nLandmarks RESET reset_nLandmarks STORED true)
     BR_PROPERTY(QList<int>, indices, QList<int>())
     BR_PROPERTY(QString, metaName, "")
     BR_PROPERTY(bool, append, false)
-    BR_PROPERTY(int, nLandmarks, 51)
     void projectMetadata(const File &src, File &dst) const
     {
         dst = src;
-        if (src.points().size() != nLandmarks) {
-            if (Globals->verbose)
-                qDebug() << "Warning: Face has " << src.points().size() << "points; should be " << nLandmarks;
-            dst.fte = true;
-            return;
-        }
-        int x1 = 0,
-            y1 = 0;
+        float x1 = 0, y1 = 0;
         for (int i = 0; i < indices.size(); i++) {
             x1 += src.points()[indices[i]].x();
@@ -202,7 +202,8 @@ class CascadeTransform : public MetaTransform
     Q_PROPERTY(int minSize READ get_minSize WRITE set_minSize RESET reset_minSize STORED false)
     Q_PROPERTY(int minNeighbors READ get_minNeighbors WRITE set_minNeighbors RESET reset_minNeighbors STORED false)
     Q_PROPERTY(bool ROCMode READ get_ROCMode WRITE set_ROCMode RESET reset_ROCMode STORED false)
-    
+    Q_PROPERTY(float scaleFactor READ get_scaleFactor WRITE set_scaleFactor RESET reset_scaleFactor STORED false)
+
     // Training parameters 
     Q_PROPERTY(int numStages READ get_numStages WRITE set_numStages RESET reset_numStages STORED false) 
     Q_PROPERTY(int w READ get_w WRITE set_w RESET reset_w STORED false)
@@ -227,7 +228,8 @@ class CascadeTransform : public MetaTransform
     BR_PROPERTY(int, minSize, 64)
     BR_PROPERTY(int, minNeighbors, 5)
     BR_PROPERTY(bool, ROCMode, false)
-        
+    BR_PROPERTY(float, scaleFactor, 1.2)
+
     // Training parameters - Default values provided trigger OpenCV defaults
     BR_PROPERTY(int, numStages, -1)
     BR_PROPERTY(int, w, -1)
@@ -410,8 +412,19 @@ class CascadeTransform : public MetaTransform
                 std::vector<Rect> rects;
                 std::vector<int> rejectLevels;
                 std::vector<double> levelWeights;
-                if (ROCMode) cascade->detectMultiScale(m, rects, rejectLevels, levelWeights, 1.2, minNeighbors, flags | CASCADE_SCALE_IMAGE, Size(minSize, minSize), Size(), true);
-                else         cascade->detectMultiScale(m, rects, 1.2, minNeighbors, flags, Size(minSize, minSize));
+                if (ROCMode) cascade->detectMultiScale(m, rects, rejectLevels, levelWeights, scaleFactor, minNeighbors, flags | CASCADE_SCALE_IMAGE, Size(minSize, minSize), Size(), true);
+                else         cascade->detectMultiScale(m, rects, scaleFactor, minNeighbors, flags, Size(minSize, minSize));
+
+                // It appears that flags is ignored for new model files:
+                // http://docs.opencv.org/modules/objdetect/doc/cascade_classification.html#cascadeclassifier-detectmultiscale
+                if ((flags == CASCADE_FIND_BIGGEST_OBJECT) && (rects.size() > 1)) {
+                    Rect biggest = rects[0];
+                    for (size_t j=0; j<rects.size(); j++)
+                        if (rects[j].area() > biggest.area())
+                            biggest = rects[j];
+                    rects.clear();
+                    rects.push_back(biggest);
+                }
                 bool empty = false;
                 if (!enrollAll && rects.empty()) {
@@ -426,7 +439,7 @@ class CascadeTransform : public MetaTransform
                         u.file.set("Confidence",-std::numeric_limits<float>::max());
                     } else if (rejectLevels.size() > j)
                         u.file.set("Confidence", rejectLevels[j]*levelWeights[j]);
-                    else 
+                    else
                         u.file.set("Confidence", rects[j].area());
                     const QRectF rect = OpenCVUtils::fromRect(rects[j]);
                     u.file.appendRect(rect);
@@ -24,28 +24,35 @@ namespace br
  * \brief Convert values of key_X, key_Y, key_Width, key_Height to a rect.
  * \author Jordan Cheney \cite JordanCheney
  */
-class KeyToRectTransform : public UntrainableMetadataTransform
+class KeyToLandmarkTransform : public UntrainableMetadataTransform
 {
     Q_OBJECT
     Q_PROPERTY(QString key READ get_key WRITE set_key RESET reset_key STORED false)
+    Q_PROPERTY(bool point READ get_point WRITE set_point RESET reset_point STORED false)
     BR_PROPERTY(QString, key, "")
+    BR_PROPERTY(bool, point, false)
     void projectMetadata(const File &src, File &dst) const
     {
         dst = src;
-        if (src.contains(QStringList() << key + "_X" << key + "_Y" << key + "_Width" << key + "_Height"))
-            dst.appendRect(QRectF(src.get<int>(key + "_X"),
-                                  src.get<int>(key + "_Y"),
-                                  src.get<int>(key + "_Width"),
-                                  src.get<int>(key + "_Height")));
+        if (point) {
+            if (src.contains(QStringList() << key + "_X" << key + "_Y"))
+                dst.appendPoint(QPointF(src.get<float>(key + "_X"),
+                                        src.get<float>(key + "_Y")));
+        } else {
+            if (src.contains(QStringList() << key + "_X" << key + "_Y" << key + "_Width" << key + "_Height"))
+                dst.appendRect(QRectF(src.get<float>(key + "_X"),
+                                      src.get<float>(key + "_Y"),
+                                      src.get<float>(key + "_Width"),
+                                      src.get<float>(key + "_Height")));
+        }
     }
-
 };
-BR_REGISTER(Transform, KeyToRectTransform)
+BR_REGISTER(Transform, KeyToLandmarkTransform)
 } // namespace br
-#include "metadata/keytorect.moc"
+#include "metadata/keytolandmark.moc"
@@ -31,15 +31,22 @@ class RemoveTemplatesTransform : public UntrainableMetaTransform
     Q_OBJECT
     Q_PROPERTY(QString regexp READ get_regexp WRITE set_regexp RESET reset_regexp STORED false)
     Q_PROPERTY(QString key READ get_key WRITE set_key RESET reset_key STORED false)
+    Q_PROPERTY(bool keep READ get_keep WRITE set_keep RESET reset_keep STORED false)
     BR_PROPERTY(QString, regexp, "")
     BR_PROPERTY(QString, key, "")
+    BR_PROPERTY(bool, keep, false)
     void project(const Template &src, Template &dst) const
     {
-        const QRegularExpression re(regexp);
-        const QRegularExpressionMatch match = re.match(key.isEmpty() ? src.file.suffix() : src.file.get<QString>(key));
-        if (match.hasMatch()) dst = Template();
-        else                  dst = src;
+        dst = src;
+        QRegExp re(regexp);
+        re.setPatternSyntax(QRegExp::Wildcard);
+        bool match = re.exactMatch(key.isEmpty() ? src.file.suffix() : src.file.get<QString>(key));
+
+        if (keep && !match)
+            dst.file.fte = true;
+        else if (!keep && match)
+            dst.file.fte = true;
     }
 };
 #include <QString>
-#include <stasm_lib.h>
 #include <stasmcascadeclassifier.h>
+#include <stasm_lib.h>
+#include <stasm.h>
 #include <opencv2/opencv.hpp>
 #include <openbr/plugins/openbr_internal.h>
@@ -56,18 +57,18 @@ BR_REGISTER(Initializer, StasmInitializer)
  * \brief Wraps STASM key point detector
  * \author Scott Klum \cite sklum
  */
-class StasmTransform : public UntrainableTransform
+class StasmTransform : public UntrainableMetaTransform
 {
     Q_OBJECT
     Q_PROPERTY(bool stasm3Format READ get_stasm3Format WRITE set_stasm3Format RESET reset_stasm3Format STORED false)
     BR_PROPERTY(bool, stasm3Format, false)
-    Q_PROPERTY(bool clearLandmarks READ get_clearLandmarks WRITE set_clearLandmarks RESET reset_clearLandmarks STORED false)
-    BR_PROPERTY(bool, clearLandmarks, false)
     Q_PROPERTY(QList<float> pinPoints READ get_pinPoints WRITE set_pinPoints RESET reset_pinPoints STORED false)
     BR_PROPERTY(QList<float>, pinPoints, QList<float>())
     Q_PROPERTY(QStringList pinLabels READ get_pinLabels WRITE set_pinLabels RESET reset_pinLabels STORED false)
     BR_PROPERTY(QStringList, pinLabels, QStringList())
+    Q_PROPERTY(QString inputVariable READ get_inputVariable WRITE set_inputVariable RESET reset_inputVariable STORED false)
+    BR_PROPERTY(QString, inputVariable, "FrontalFace")
     Resource<StasmCascadeClassifier> stasmCascadeResource;
@@ -77,86 +78,119 @@ class StasmTransform : public UntrainableTransform
         stasmCascadeResource.setResourceMaker(new StasmResourceMaker());
     }
-    void project(const Template &src, Template &dst) const
+    QList<QPointF> convertLandmarks(int nLandmarks, float *landmarks) const
     {
-        Mat stasmSrc(src);
-        if (src.m().channels() == 3)
-            cvtColor(src, stasmSrc, CV_BGR2GRAY);
-        else if (src.m().channels() != 1)
-            qFatal("Stasm expects single channel matrices.");
-
-        if (!stasmSrc.isContinuous())
-            qFatal("Stasm expects continuous matrix data.");
-        dst = src;
-
-        int foundFace = 0;
-        int nLandmarks = stasm_NLANDMARKS;
-        float landmarks[2 * stasm_NLANDMARKS];
-
-        bool searchPinned = false;
-
-        QPointF rightEye, leftEye;
-        /* Two use cases are accounted for:
-         * 1. Pin eyes without normalization: in this case the string list should contain the KEYS for right then left eyes, respectively.
-         * 2. Pin eyes with normalization: in this case the string list should contain the COORDINATES of the right then left eyes, respectively.
-         * Currently, we only support normalization with a transformation such that the src file contains Affine_0 and Affine_1.  Checking for
-         * these keys prevents us from pinning eyes on a face that wasn't actually transformed (see AffineTransform).
-         * If both cases fail, we default to stasm_search_single. */
-
-        if (!pinPoints.isEmpty() && src.file.contains("Affine_0") && src.file.contains("Affine_1")) {
-            rightEye = QPointF(pinPoints.at(0), pinPoints.at(1));
-            leftEye = QPointF(pinPoints.at(2), pinPoints.at(3));
-            searchPinned = true;
-        } else if (!pinLabels.isEmpty()) {
-            rightEye = src.file.get<QPointF>(pinLabels.at(0), QPointF());
-            leftEye = src.file.get<QPointF>(pinLabels.at(1), QPointF());
-            searchPinned = true;
-        }
-	
-        if (searchPinned) {
-            float pins[2 * stasm_NLANDMARKS];
-
-            for (int i = 0; i < nLandmarks; i++) {
-                if      (i == 38) /* Stasm Right Eye */ { pins[2*i] = rightEye.x(); pins[2*i+1] = rightEye.y(); }
-                else if (i == 39) /* Stasm Left Eye  */ { pins[2*i] = leftEye.x();  pins[2*i+1] = leftEye.y(); }
-                else { pins[2*i] = 0; pins[2*i+1] = 0; }
-            }
+        if (stasm3Format)
+            stasm_convert_shape(landmarks, 76);
-            stasm_search_pinned(landmarks, pins, reinterpret_cast<const char*>(stasmSrc.data), stasmSrc.cols, stasmSrc.rows, NULL);
-
-            // The ASM in Stasm is guaranteed to converge in this case
-            foundFace = 1;
+        QList<QPointF> points;
+        for (int i = 0; i < nLandmarks; i++) {
+            QPointF point(landmarks[2 * i], landmarks[2 * i + 1]);
+            points.append(point);
         }
-        if (!foundFace) {
-            StasmCascadeClassifier *stasmCascade = stasmCascadeResource.acquire();
-            stasm_search_single(&foundFace, landmarks, reinterpret_cast<const char*>(stasmSrc.data), stasmSrc.cols, stasmSrc.rows, *stasmCascade, NULL, NULL);
-            stasmCascadeResource.release(stasmCascade);
-        }
+        return points;
+    }
-        if (stasm3Format) {
-            nLandmarks = 76;
-            stasm_convert_shape(landmarks, nLandmarks);
-        }
+    void project(const Template &src, Template &dst) const
+    {
+        TemplateList temp;
+        project(TemplateList() << src, temp);
+        if (!temp.isEmpty()) dst = temp.first();
+    }
-        // For convenience, if these are the only points/rects we want to deal with as the algorithm progresses
-        if (clearLandmarks) {
-            dst.file.clearPoints();
-            dst.file.clearRects();
-        }
+    void project(const TemplateList &src, TemplateList &dst) const
+    {
+        foreach (const Template &t, src) {
+            Mat stasmSrc(t);
+            if (t.m().channels() == 3)
+                cvtColor(t, stasmSrc, CV_BGR2GRAY);
+            else if (t.m().channels() != 1)
+                qFatal("Stasm expects single channel matrices.");
+
+            if (!stasmSrc.isContinuous())
+                qFatal("Stasm expects continuous matrix data.");
+
+            int foundFace = 0;
+            int nLandmarks = stasm_NLANDMARKS;
+            float landmarks[2 * stasm_NLANDMARKS];
+
+            bool searchPinned = false;
+
+            QPointF rightEye, leftEye;
+            /* Two use cases are accounted for:
+             * 1. Pin eyes without normalization: in this case the string list should contain the KEYS for right then left eyes, respectively.
+             * 2. Pin eyes with normalization: in this case the string list should contain the COORDINATES of the right then left eyes, respectively.
+             * Currently, we only support normalization with a transformation such that the src file contains Affine_0 and Affine_1.  Checking for
+             * these keys prevents us from pinning eyes on a face that wasn't actually transformed (see AffineTransform).
+             * If both cases fail, we default to stasm_search_single. */
+
+            if (!pinPoints.isEmpty() && t.file.contains("Affine_0") && t.file.contains("Affine_1")) {
+                rightEye = QPointF(pinPoints.at(0), pinPoints.at(1));
+                leftEye = QPointF(pinPoints.at(2), pinPoints.at(3));
+                searchPinned = true;
+            } else if (!pinLabels.isEmpty()) {
+                rightEye = t.file.get<QPointF>(pinLabels.at(0), QPointF());
+                leftEye = t.file.get<QPointF>(pinLabels.at(1), QPointF());
+                searchPinned = true;
+            }
+
+            if (searchPinned) {
+                float pins[2 * stasm_NLANDMARKS];
+
+                for (int i = 0; i < nLandmarks; i++) {
+                    if      (i == 38) /* Stasm Right Eye */ { pins[2*i] = rightEye.x(); pins[2*i+1] = rightEye.y(); }
+                    else if (i == 39) /* Stasm Left Eye  */ { pins[2*i] = leftEye.x();  pins[2*i+1] = leftEye.y(); }
+                    else { pins[2*i] = 0; pins[2*i+1] = 0; }
+                }
+
+                stasm_search_pinned(landmarks, pins, reinterpret_cast<const char*>(stasmSrc.data), stasmSrc.cols, stasmSrc.rows, NULL);
+
+                // The ASM in Stasm is guaranteed to converge in this case
+                foundFace = 1;
+                Template u(t.file, t.m());
+                QList<QPointF> points = convertLandmarks(nLandmarks, landmarks);
+                u.file.set("StasmRightEye", points[38]);
+                u.file.set("StasmLeftEye", points[39]);
+                u.file.appendPoints(points);
+                dst.append(u);
+            }
-        if (!foundFace) {
-            if (Globals->verbose) qWarning("No face found in %s.", qPrintable(src.file.fileName()));
-            dst.file.fte = true;
-        } else {
-            QList<QPointF> points;
-            for (int i = 0; i < nLandmarks; i++) {
-                QPointF point(landmarks[2 * i], landmarks[2 * i + 1]);
-                points.append(point);
+            if (!foundFace) {
+                StasmCascadeClassifier *stasmCascade = stasmCascadeResource.acquire();
+                foundFace = 1;
+                stasm::FaceDet detection;
+
+                QList<QRectF> rects = t.file.contains(inputVariable) ? QList<QRectF>() << t.file.get<QRectF>(inputVariable) : t.file.rects();
+                if (!rects.isEmpty()) detection.iface_ = 0;
+                for (int i=0; i<rects.size(); i++) {
+                    Rect rect = OpenCVUtils::toRect(rects[i]);
+                    stasm::DetPar detpar;
+                    detpar.x = rect.x + rect.width / 2.;
+                    detpar.y = rect.y + rect.height / 2.;
+                    detpar.width  = double(rect.width);
+                    detpar.height = double(rect.height);
+                    detpar.yaw = 0;
+                    detpar.eyaw = stasm::EYAW00;
+                    detection.detpars_.push_back(detpar);
+                }
+
+                while (foundFace) {
+                    stasm_search_auto(&foundFace, landmarks, reinterpret_cast<const char*>(stasmSrc.data), stasmSrc.cols, stasmSrc.rows, *stasmCascade, detection);
+                    if (foundFace) {
+                        Template u(t.file, t.m());
+                        QList<QPointF> points = convertLandmarks(nLandmarks, landmarks);
+                        u.file.set("StasmRightEye", points[38]);
+                        u.file.set("StasmLeftEye", points[39]);
+                        u.file.appendPoints(points);
+                        dst.append(u);
+                    }
+
+                    if (!Globals->enrollAll)
+                        break;
+                }
+                stasmCascadeResource.release(stasmCascade);
             }
-            dst.file.set("StasmRightEye", points[38]);
-            dst.file.set("StasmLeftEye", points[39]);
-            dst.file.appendPoints(points);
         }
     }
 };
@@ -6,44 +6,80 @@
 namespace br
 {
+typedef QPair<float,int> Pair;
+
 /*!
  * \ingroup outputs
  * \brief Outputs the k-Nearest Neighbors from the gallery for each probe.
  * \author Ben Klein \cite bhklein
  */
-class knnOutput : public MatrixOutput
+class knnOutput : public Output
 {
     Q_OBJECT
+    int rowBlock, columnBlock;
+    size_t headerSize, k;
+    cv::Mat blockScores;
+
     ~knnOutput()
     {
-        size_t num_probes = (size_t)queryFiles.size();
-        if (targetFiles.isEmpty() || queryFiles.isEmpty()) return;
-        size_t k = file.get<size_t>("k", 20);
+        writeBlock();
+    }
-        if ((size_t)targetFiles.size() < k)
-            qFatal("Gallery size %s is smaller than k = %s.", qPrintable(QString::number(targetFiles.size())), qPrintable(QString::number(k)));
+    void setBlock(int rowBlock, int columnBlock)
+    {
+        if ((rowBlock == 0) && (columnBlock == 0)) {
+            k = file.get<size_t>("k", 20);
+            QFile f(file);
+            if (!f.open(QFile::WriteOnly))
+                qFatal("Unable to open %s for writing.", qPrintable(file));
+            size_t querySize = (size_t)queryFiles.size();
+            f.write((const char*) &querySize, sizeof(size_t));
+            f.write((const char*) &k, sizeof(size_t));
+            headerSize = 2 * sizeof(size_t);
+        } else {
+            writeBlock();
+        }
-        QFile f(file);
-        if (!f.open(QFile::WriteOnly))
-            qFatal("Unable to open %s for writing.", qPrintable(file));
-        f.write((const char*) &num_probes, sizeof(size_t));
-        f.write((const char*) &k, sizeof(size_t));
+        this->rowBlock = rowBlock;
+        this->columnBlock = columnBlock;
+
+        int matrixRows  = std::min(queryFiles.size()-rowBlock*this->blockRows, blockRows);
+        int matrixCols  = std::min(targetFiles.size()-columnBlock*this->blockCols, blockCols);
+
+        blockScores = cv::Mat(matrixRows, matrixCols, CV_32FC1);
+    }
-        QVector<Candidate> neighbors; neighbors.reserve(num_probes*k);
+    void setRelative(float value, int i, int j)
+    {
+        blockScores.at<float>(i,j) = value;
+    }
+
+    void set(float value, int i, int j)
+    {
+        (void) value; (void) i; (void) j;
+        qFatal("Logic error.");
+    }
+
+    void writeBlock()
+    {
+        QFile f(file);
+        if (!f.open(QFile::ReadWrite))
+            qFatal("Unable to open %s for modifying.", qPrintable(file));
+        QVector<Candidate> neighbors; neighbors.reserve(k * blockScores.rows);
-        for (size_t i=0; i<num_probes; i++) {
-            typedef QPair<float,int> Pair;
+        for (int i=0; i<blockScores.rows; i++) {
             size_t rank = 0;
-            foreach (const Pair &pair, Common::Sort(OpenCVUtils::matrixToVector<float>(data.row(i)), true)) {
-                if (QString(targetFiles[pair.second]) != QString(queryFiles[i])) {
+            foreach (const Pair &pair, Common::Sort(OpenCVUtils::matrixToVector<float>(blockScores.row(i)), true)) {
+                if (QString(targetFiles[pair.second]) != QString(queryFiles[rowBlock*this->blockRows+i])) {
                     Candidate candidate((size_t)pair.second, pair.first);
                     neighbors.push_back(candidate);
                     if (++rank >= k) break;
                 }
             }
         }
-        f.write((const char*) neighbors.data(), num_probes * k * sizeof(Candidate));
+        f.seek(headerSize + sizeof(Candidate)*quint64(rowBlock*this->blockRows)*k);
+        f.write((const char*) neighbors.data(), blockScores.rows * k * sizeof(Candidate));
         f.close();
     }
 };
@@ -80,23 +80,25 @@ class HaarRepresentation : public Representation
         }
     }
-    void preprocess(const Mat &src, Mat &dst) const
+    Template preprocess(const Template &src) const
     {
+        Template dst;
         integral(src, dst);
+        return dst;
     }
-    float evaluate(const Mat &image, int idx) const
+    float evaluate(const Template &src, int idx) const
     {
-        return (float)features[idx].calc(image);
+        return (float)features[idx].calc(src.m());
     }
-    Mat evaluate(const Mat &image, const QList<int> &indices) const
+    Mat evaluate(const Template &src, const QList<int> &indices) const
     {
         int size = indices.empty() ? numFeatures() : indices.size();
         Mat result(1, size, CV_32FC1);
         for (int i = 0; i < size; i++)
-            result.at<float>(i) = evaluate(image, indices.empty() ? i : indices[i]);
+            result.at<float>(i) = evaluate(src, indices.empty() ? i : indices[i]);
         return result;
     }
@@ -39,32 +39,36 @@ class MBLBPRepresentation : public Representation
     void init()
     {
-        int offset = winWidth + 1;
-        for (int x = 0; x < winWidth; x++ )
-            for (int y = 0; y < winHeight; y++ )
-                for (int w = 1; w <= winWidth / 3; w++ )
-                    for (int h = 1; h <= winHeight / 3; h++ )
-                        if ((x+3*w <= winWidth) && (y+3*h <= winHeight) )
-                            features.append(Feature(offset, x, y, w, h ) );
+        if (features.isEmpty()) {
+            int offset = winWidth + 1;
+            for (int x = 0; x < winWidth; x++ )
+                for (int y = 0; y < winHeight; y++ )
+                    for (int w = 1; w <= winWidth / 3; w++ )
+                        for (int h = 1; h <= winHeight / 3; h++ )
+                            if ((x+3*w <= winWidth) && (y+3*h <= winHeight) )
+                                features.append(Feature(offset, x, y, w, h ) );
+        }
     }
-    void preprocess(const Mat &src, Mat &dst) const
+    Template preprocess(const Template &src) const
     {
+        Template dst;
         integral(src, dst);
+        return dst;
     }
-    float evaluate(const Mat &image, int idx) const
+    float evaluate(const Template &src, int idx) const
     {
-        return (float)features[idx].calc(image);
+        return (float)features[idx].calc(src.m());
     }
-    Mat evaluate(const Mat &image, const QList<int> &indices) const
+    Mat evaluate(const Template &src, const QList<int> &indices) const
     {
         int size = indices.empty() ? numFeatures() : indices.size();
         Mat result(1, size, CV_32FC1);
         for (int i = 0; i < size; i++)
-            result.at<float>(i) = evaluate(image, indices.empty() ? i : indices[i]);
+            result.at<float>(i) = evaluate(src, indices.empty() ? i : indices[i]);
         return result;
     }
@@ -16,23 +16,25 @@ class NPDRepresentation : public Representation
     void init()
     {
-        for (int p1 = 0; p1 < (winWidth * winHeight); p1++)
-            for (int p2 = p1; p2 < (winWidth * winHeight); p2++)
-                features.append(Feature(p1, p2));
+        if (features.isEmpty()) {
+            for (int p1 = 0; p1 < (winWidth * winHeight); p1++)
+                for (int p2 = p1; p2 < (winWidth * winHeight); p2++)
+                    features.append(Feature(p1, p2));
+        }
     }
-    float evaluate(const Mat &image, int idx) const
+    float evaluate(const Template &src, int idx) const
     {
-        return features[idx].calc(image);
+        return features[idx].calc(src.m());
     }
-    Mat evaluate(const Mat &image, const QList<int> &indices) const
+    Mat evaluate(const Template &src, const QList<int> &indices) const
     {
         int size = indices.empty() ? numFeatures() : indices.size();
         Mat result(1, size, CV_32FC1);
         for (int i = 0; i < size; i++)
-            result.at<float>(i) = evaluate(image, indices.empty() ? i : indices[i]);
+            result.at<float>(i) = evaluate(src, indices.empty() ? i : indices[i]);
         return result;
     }
@@ -49,10 +51,10 @@ class NPDRepresentation : public Representation
     struct Feature
     {
         Feature() {}
-        Feature(int p1, int p2) { p[0] = p1; p[1] = p2; }
+        Feature(int p0, int p1) : p0(p0), p1(p1) {}
         float calc(const Mat &image) const;
-        int p[2];
+        int p0, p1;
     };
     QList<Feature> features;
 };
@@ -61,9 +63,9 @@ BR_REGISTER(Representation, NPDRepresentation)
 inline float NPDRepresentation::Feature::calc(const Mat &image) const
 {
-    const int *ptr = image.ptr<int>();
-    int v1 = ptr[p[0]], v2 = ptr[p[1]];
-    return v1 == 0 && v2 == 0 ? 0 : ((float)(v1 - v2)) / (v1 + v2);
+    const uchar *ptr = image.ptr();
+    int v1 = (int)ptr[p0], v2 = (int)ptr[p1];
+    return (v1 + v2) == 0 ? 0 : (1.0 * (v1 - v2)) / (v1 + v2);
 }
 } // namespace br
@@ -28,9 +28,9 @@ class RandomRepresentation : public Representation
     QList<int> features;
-    void train(const QList<Mat> &images, const QList<float> &labels)
+    void train(const TemplateList &data)
     {
-        representation->train(images, labels);
+        representation->train(data);
         const int nFeatures = representation->numFeatures();
@@ -40,17 +40,17 @@ class RandomRepresentation : public Representation
         features = Common::RandSample(count,nFeatures,0,true);
     }
-    void preprocess(const Mat &src, Mat &dst) const
+    Template preprocess(const Template &src) const
     {
-        representation->preprocess(src,dst);
+        return representation->preprocess(src);
     }
-    float evaluate(const Mat &image, int idx) const
+    float evaluate(const Template &src, int idx) const
     {
-        return representation->evaluate(image,features[idx]);
+        return representation->evaluate(src, features[idx]);
     }
-    Mat evaluate(const Mat &image, const QList<int> &indices) const
+    Mat evaluate(const Template &src, const QList<int> &indices) const
     {
         QList<int> newIndices;
         if (indices.empty())
@@ -59,7 +59,7 @@ class RandomRepresentation : public Representation
             for (int i = 0; i < indices.size(); i++)
                 newIndices.append(features[indices[i]]);
-        return representation->evaluate(image,newIndices);
+        return representation->evaluate(src, newIndices);
     }
     int numFeatures() const
@@ -108,6 +108,3 @@ BR_REGISTER(Representation, RandomRepresentation)
 } // namespace br
 #include "representation/random.moc"
-
-
-
 #!/usr/bin/env python
-from PIL import Image
 import csv, sys, json, argparse
+from brpy.html_viz import crop_to_bb
 parser = argparse.ArgumentParser(description='Visualize face cluster results in an HTML page.')
 parser.add_argument('input_file', type=str, help='Results from clustering (in csv format)')
@@ -16,24 +16,15 @@ clustmap = dict()
 with open(args.input_file) as f:
     for line in csv.DictReader(f):
         c = int(line[args.cluster_key])
+        if c not in clustmap:
+            clustmap[c] = []
         x,y,width,height = [ float(line[k]) for k in ('Face_X','Face_Y','Face_Width','Face_Height') ]
         imname = '%s/%s' % (args.img_loc, line['File'])
         try:
-            img = Image.open(imname)
-            imwidth, imheight = img.size
+            html = crop_to_bb(x,y,width,height,imname,maxheight=400)
         except IOError:
             print('problem with %s' % imname)
             continue
-        ratio = maxheight / height
-        # note for future me:
-        # image is cropped with div width/height + overflow:hidden,
-        # resized with img height,
-        # and positioned with img margin
-        html = '<div style="overflow:hidden;display:inline-block;width:%ipx;height:%ipx;">' % (width*ratio, maxheight)
-        html += '<img src="%s" style="height:%ipx;margin:-%ipx 0 0 -%ipx;"/>' % (imname, imheight*ratio, y*ratio, x*ratio)
-        html += '</div>'
-        if c not in clustmap:
-            clustmap[c] = []
         clustmap[c].append(html)
 # browsers crash for a DOM with this many img tags,
+'''
+Some funcs to generate HTML visualizations.
+Run from the folder you intend to save the HTML page so
+the relative paths in the HTML file are correct and PIL can find the images on disk.
+Requires local images, but should be pretty easy to set up an apache server (or whatev)
+and host them as long as the relative paths remain the same on ya serva.
+'''
+
+from PIL import Image
+
+def crop_to_bb(x, y, width, height, imname, maxheight=None):
+    '''
+    Generates an HTML string that crops to a given bounding box and resizes to maxheight pixels.
+    A maxheight of None will keep the original size (default).
+    When two crops are put next to each other, they will be inline. To make each crop its own line, wrap it in a div.
+    '''
+    img = Image.open(imname)
+    imwidth, imheight = img.size
+    if not maxheight:
+        maxheight = height
+    ratio = maxheight / height
+    # note for future me:
+    # image is cropped with div width/height + overflow:hidden,
+    # resized with img height,
+    # and positioned with img margin
+    html = '<div style="overflow:hidden; display:inline-block; width:%ipx; height:%ipx;">' % (width*ratio, maxheight)
+    html += '<img src="%s" style="height:%ipx; margin:-%ipx 0 0 -%ipx;"/>' % (imname, imheight*ratio, y*ratio, x*ratio)
+    html += '</div>'
+    return html
+
+def bbs_for_image(imname, bbs, maxheight=None, colors=None):
+    '''
+    Generates an HTML string for an image with bounding boxes.
+    bbs: iterable of (x,y,width,height) bounding box tuples
+    '''
+    img = Image.open(imname)
+    imwidth, imheight = img.size
+    if not maxheight:
+        maxheight = imheight
+    ratio = maxheight/imheight
+    html = [
+            '<div style="position:relative">',
+            '<img src="%s" style="height:%ipx" />' % (imname, maxheight)
+           ]
+    if not colors:
+        colors = ['green']*len(bbs)
+    html.extend([ bb(*box, ratio=ratio, color=color) for color,box in zip(colors,bbs) ])
+    html.append('</div>')
+    return '\n'.join(html)
+
+
+def bb(x, y, width, height, ratio=1.0, color='green'):
+    '''
+    Generates an HTML string bounding box.
+    '''
+    html = '<div style="position:absolute; border:2px solid %s; color:%s; left:%ipx; top:%ipx; width:%ipx; height:%ipx;"></div>'
+    return html % (color, color, x*ratio, y*ratio, width*ratio, height*ratio)
@@ -86,12 +86,14 @@ endfunction()
 function(install_qt_platforms)
   if(${BR_INSTALL_DEPENDENCIES})
     if(CMAKE_HOST_WIN32)
-      #TODO
+      install(FILES ${_qt5Core_install_prefix}/plugins/platforms/qwindows.dll
+              DESTINATION bin/platforms)
     elseif(CMAKE_HOST_APPLE)
       install(FILES ${_qt5Core_install_prefix}/plugins/platforms/libqcocoa.dylib
               DESTINATION bin/platforms)
     else()
-      #TODO
+      install(FILES ${_qt5Core_install_prefix}/plugins/platforms/libqlinuxfb.so
+              DESTINATION bin/platforms)
     endif()
   endif()
 endfunction()
@@ -217,6 +217,12 @@ formatData &lt;- function(type=&quot;eval&quot;) {
         NormLength <<- data[grep("NormLength",data$Plot),-c(1)]
         sample <<- readImageData(Sample)
         rows <<- sample[[1]]$value
+    } else if (type == "knn") {
+        # Split data into individual plots
+        IET <<- data[grep("IET",data$Plot),-c(1)]
+        IET$Y <<- as.numeric(as.character(IET$Y))
+        CMC <<- data[grep("CMC",data$Plot),-c(1)]
+        CMC$Y <<- as.numeric(as.character(CMC$Y))
     }
 }