Reverted 2 commits to avoid regression. Added first attempt at Haar representation

Jordan Cheney
1 parent 9eaaf62b
Showing 6 changed files with 277 additions and 143 deletions
openbr/core/boost.cpp
openbr/core/boost.h
openbr/plugins/classification/boostedforest.cpp
openbr/plugins/classification/cascade.cpp
openbr/plugins/representation/haar.cpp
openbr/plugins/representation/mblbp.cpp
@@ -154,11 +154,6 @@ void FeatureEvaluator::setImage(const Mat &amp;img, uchar clsLabel, int idx)
     representation->preprocess(img, integralImg);
 }
-void FeatureEvaluator::writeFeatures(FileStorage &fs, const Mat &featureMap) const
-{
-    representation->write(fs, featureMap);
-}
-
 //----------------------------- CascadeBoostParams -------------------------------------------------
 CascadeBoostParams::CascadeBoostParams() : minHitRate( 0.995F), maxFalseAlarm( 0.5F )
@@ -178,21 +173,6 @@ CascadeBoostParams::CascadeBoostParams( int _boostType,
     use_surrogates = use_1se_rule = truncate_pruned_tree = false;
 }
-void CascadeBoostParams::write( FileStorage &fs ) const
-{
-    string boostTypeStr = boost_type == CvBoost::DISCRETE ? CC_DISCRETE_BOOST :
-                          boost_type == CvBoost::REAL ? CC_REAL_BOOST :
-                          boost_type == CvBoost::LOGIT ? CC_LOGIT_BOOST :
-                          boost_type == CvBoost::GENTLE ? CC_GENTLE_BOOST : string();
-    CV_Assert( !boostTypeStr.empty() );
-    fs << CC_BOOST_TYPE << boostTypeStr;
-    fs << CC_MINHITRATE << minHitRate;
-    fs << CC_MAXFALSEALARM << maxFalseAlarm;
-    fs << CC_TRIM_RATE << weight_trim_rate;
-    fs << CC_MAX_DEPTH << max_depth;
-    fs << CC_WEAK_COUNT << weak_count;
-}
-
 //---------------------------- CascadeBoostTrainData -----------------------------
 CvDTreeNode* CascadeBoostTrainData::subsample_data( const CvMat* _subsample_idx )
@@ -826,7 +806,7 @@ CvDTreeNode* CascadeBoostTree::predict( int sampleIdx ) const
     return node;
 }
-/*
+
 static void writeRecursive(FileStorage &fs, CvDTreeNode *node, int maxCatCount)
 {
     bool hasChildren = node->left ? true : false;
@@ -835,7 +815,7 @@ static void writeRecursive(FileStorage &amp;fs, CvDTreeNode *node, int maxCatCount)
     if (!hasChildren) // Write the leaf value
         fs << "value" << node->value; // value of the node. Only relevant for leaf nodes
     else { // Write the splitting information and then the children
-        if (maxCatCount > 0) {
+        if (maxCatCount > 1) {
             fs << "subset" << "[:";
             for (int i = 0; i < ((maxCatCount + 31) / 32); i++)
                 fs << node->split->subset[i]; // subset to split on (categorical features)
@@ -857,7 +837,7 @@ void CascadeBoostTree::write(FileStorage &amp;fs)
     writeRecursive(fs, root, ((CascadeBoostTrainData*)data)->featureEvaluator->getMaxCatCount());
     fs << "}";
 }
-
+/*
 static void readRecursive(const FileNode &fn, CvDTreeNode *node, CvDTreeTrainData *data)
 {
     bool hasChildren = (int)fn["hasChildren"];
@@ -897,67 +877,8 @@ void CascadeBoostTree::read(const FileNode &amp;fn, CvBoost* _ensemble, CvDTreeTrain
     root = data->new_node(0, 0, 0, 0);
     readRecursive(fn, root, data);
-}*/
-
-void CascadeBoostTree::write(FileStorage &fs)
-{
-    int maxCatCount = ((CascadeBoostTrainData*)data)->featureEvaluator->getMaxCatCount();
-    int subsetN = (maxCatCount + 31)/32;
-    queue<CvDTreeNode*> internalNodesQueue;
-    int size = (int)pow( 2.f, (float)ensemble->get_params().max_depth);
-    Ptr<float> leafVals = new float[size];
-    int leafValIdx = 0;
-    int internalNodeIdx = 1;
-    CvDTreeNode* tempNode;
-
-    CV_DbgAssert( root );
-    internalNodesQueue.push( root );
-
-    fs << "{";
-    fs << CC_INTERNAL_NODES << "[:";
-    while (!internalNodesQueue.empty())
-    {
-        tempNode = internalNodesQueue.front();
-        CV_Assert( tempNode->left );
-        if ( !tempNode->left->left && !tempNode->left->right) // left node is leaf
-        {
-            leafVals[-leafValIdx] = (float)tempNode->left->value;
-            fs << leafValIdx-- ;
-        }
-        else
-        {
-            internalNodesQueue.push( tempNode->left );
-            fs << internalNodeIdx++;
-        }
-        CV_Assert( tempNode->right );
-        if ( !tempNode->right->left && !tempNode->right->right) // right node is leaf
-        {
-            leafVals[-leafValIdx] = (float)tempNode->right->value;
-            fs << leafValIdx--;
-        }
-        else
-        {
-            internalNodesQueue.push( tempNode->right );
-            fs << internalNodeIdx++;
-        }
-        int fidx = tempNode->split->var_idx;
-
-        fs << fidx;
-        if ( !maxCatCount )
-            fs << tempNode->split->ord.c;
-        else
-            for( int i = 0; i < subsetN; i++ )
-                fs << tempNode->split->subset[i];
-        internalNodesQueue.pop();
-    }
-    fs << "]"; // CC_INTERNAL_NODES
-
-    fs << CC_LEAF_VALUES << "[:";
-    for (int ni = 0; ni < -leafValIdx; ni++)
-        fs << leafVals[ni];
-    fs << "]"; // CC_LEAF_VALUES
-    fs << "}";
 }
+*/
 void CascadeBoostTree::split_node_data( CvDTreeNode* node )
 {
@@ -1214,6 +1135,7 @@ bool CascadeBoost::train( const FeatureEvaluator* _featureEvaluator,
             break;
         }
+        classifiers.append(tree);
         cvSeqPush( weak, &tree );
         update_weights( tree );
         trim_weights();
@@ -1542,20 +1464,3 @@ bool CascadeBoost::isErrDesired()
     return falseAlarm <= maxFalseAlarm;
 }
-void CascadeBoost::write(FileStorage &fs) const
-{
-//    char cmnt[30];
-    CascadeBoostTree* weakTree;
-    fs << CC_WEAK_COUNT << weak->total;
-    fs << CC_STAGE_THRESHOLD << threshold;
-    fs << CC_WEAK_CLASSIFIERS << "[";
-    for( int wi = 0; wi < weak->total; wi++)
-    {
-        /*sprintf( cmnt, "tree %i", wi );
-        cvWriteComment( fs, cmnt, 0 );*/
-        weakTree = *((CascadeBoostTree**) cvGetSeqElem( weak, wi ));
-        weakTree->write(fs);
-    }
-    fs << "]";
-}
-
@@ -55,7 +55,6 @@ struct FeatureEvaluator
     ~FeatureEvaluator() {}
     void init(Representation *_representation, int _maxSampleCount);
     void setImage(const cv::Mat& img, uchar clsLabel, int idx);
-    void writeFeatures(cv::FileStorage &fs, const cv::Mat& featureMap) const;
     float operator()(int featureIdx, int sampleIdx) const { return representation->evaluate(data.row(sampleIdx), featureIdx); }
     int getNumFeatures() const { return representation->numFeatures(); }
@@ -77,7 +76,6 @@ struct CascadeBoostParams : CvBoostParams
     CascadeBoostParams(int _boostType, float _minHitRate, float _maxFalseAlarm,
                        double _weightTrimRate, int _maxDepth, int _maxWeakCount);
     virtual ~CascadeBoostParams() {}
-    void write( cv::FileStorage &fs ) const;
 };
 struct CascadeBoostTrainData : CvDTreeTrainData
@@ -128,13 +126,15 @@ public:
     virtual float predict( int sampleIdx, bool returnSum = false ) const;
     float getThreshold() const { return threshold; }
-    void write(cv::FileStorage &fs) const;
+    QList<CvBoostTree*> getClassifiers() const { return classifiers; }
 protected:
     virtual bool set_params(const CvBoostParams& _params);
     virtual void update_weights(CvBoostTree* tree);
     virtual bool isErrDesired();
+    QList<CvBoostTree*> classifiers;
+
     float threshold;
     float minHitRate, maxFalseAlarm;
 };
@@ -6,6 +6,64 @@ using namespace cv;
 namespace br
 {
+struct Node
+{
+    Node() : left(NULL), right(NULL) {}
+
+    float value;
+
+    float threshold; // For ordered features
+    QList<int> subset; // For categorical features
+    int featureIdx;
+
+    Node *left;
+    Node *right;
+};
+
+static void buildTreeRecursive(Node *node, const CvDTreeNode *cv_node, int maxCatCount)
+{
+    if (!cv_node->left) // Write the leaf value
+        node->value = cv_node->value; // value of the node. Only relevant for leaf nodes
+    else { // Write the splitting information and then the children
+        if (maxCatCount > 1)
+            for (int i = 0; i < ((maxCatCount + 31) / 32); i++)
+                node->subset.append(cv_node->split->subset[i]); // subset to split on (categorical features)
+        else
+            node->threshold = cv_node->split->ord.c; // threshold to split on (ordered features)
+
+        node->featureIdx = cv_node->split->var_idx; // feature idx of node
+
+        node->left = new Node;
+        buildTreeRecursive(node->left, cv_node->left, maxCatCount);
+        node->right = new Node;
+        buildTreeRecursive(node->right, cv_node->right, maxCatCount);
+    }
+}
+
+static void writeRecursive(FileStorage &fs, const Node *node, int maxCatCount)
+{
+    bool hasChildren = node->left ? true : false;
+    fs << "hasChildren" << hasChildren;
+
+    if (!hasChildren) // Write the leaf value
+        fs << "value" << node->value; // value of the node. Only relevant for leaf nodes
+    else { // Write the splitting information and then the children
+        if (maxCatCount > 1) {
+            fs << "subset" << "[:";
+            for (int i = 0; i < ((maxCatCount + 31) / 32); i++)
+                fs << node->subset[i]; // subset to split on (categorical features)
+            fs << "]";
+        } else {
+            fs << "threshold" << node->threshold; // threshold to split on (ordered features)
+        }
+
+        fs << "feature_idx" << node->featureIdx; // feature idx of node
+
+        fs << "left" << "{"; writeRecursive(fs, node->left, maxCatCount); fs << "}"; // write left child
+        fs << "right" << "{"; writeRecursive(fs, node->right, maxCatCount); fs << "}"; // write right child
+    }
+}
+
 class BoostedForestClassifier : public Classifier
 {
     Q_OBJECT
@@ -24,27 +82,49 @@ class BoostedForestClassifier : public Classifier
     BR_PROPERTY(int, maxDepth, 1)
     BR_PROPERTY(int, maxWeakCount, 100)
-    CascadeBoost *boost;
-    FeatureEvaluator *featureEvaluator;
+    QList<Node*> weakClassifiers;
+    float threshold;
     void train(const QList<Mat> &images, const QList<float> &labels)
     {
         CascadeBoostParams params(CvBoost::GENTLE, minTAR, maxFAR, trimRate, maxDepth, maxWeakCount);
-        featureEvaluator = new FeatureEvaluator;
-        featureEvaluator->init(representation, images.size());
+        FeatureEvaluator featureEvaluator;
+        featureEvaluator.init(representation, images.size());
         for (int i = 0; i < images.size(); i++)
-            featureEvaluator->setImage(images[i], labels[i], i);
+            featureEvaluator.setImage(images[i], labels[i], i);
+
+        CascadeBoost boost;
+        boost.train(&featureEvaluator, images.size(), 2048, 2048, params);
+
+        threshold = boost.getThreshold();
-        boost = new CascadeBoost;
-        boost->train(featureEvaluator, images.size(), 1024, 1024, params);
+        foreach (const CvBoostTree *tree, boost.getClassifiers()) {
+            Node *root = new Node;
+            buildTreeRecursive(root, tree->get_root(), representation->maxCatCount());
+            weakClassifiers.append(root);
+        }
     }
     float classify(const Mat &image) const
     {
-        featureEvaluator->setImage(image, 0, 0);
-        return boost->predict(0);
+        float sum = 0;
+        foreach (const Node *root, weakClassifiers) {
+            const Node *node = root;
+
+            while (node->left) {
+                if (representation->maxCatCount() > 1) {
+                    int c = (int)representation->evaluate(image, node->featureIdx);
+                    node = (node->subset[c >> 5] & (1 << (c & 31))) ? node->left : node->right;
+                } else {
+                    float val = representation->evaluate(image, node->featureIdx);
+                    node = val < node->threshold ? node->left : node->right;
+                }
+            }
+            sum += node->value;
+        }
+        return sum < threshold - FLT_EPSILON ? 0.0 : 1.0;
     }
     int numFeatures() const
@@ -64,7 +144,15 @@ class BoostedForestClassifier : public Classifier
     void write(FileStorage &fs) const
     {
-        boost->write(fs);
+        fs << "weakCount" << weakClassifiers.size();
+        fs << "stageThreshold" << threshold;
+        fs << "weakClassifiers" << "[";
+        foreach (const Node *root, weakClassifiers) {
+            fs << "{";
+            writeRecursive(fs, root, representation->maxCatCount());
+            fs << "}";
+        }
+        fs << "]";
     }
 };
@@ -178,22 +178,9 @@ class CascadeClassifier : public Classifier
     void write(FileStorage &fs) const
     {
-        fs << CC_STAGE_TYPE << CC_BOOST;
-        fs << CC_FEATURE_TYPE << CC_LBP;
-        fs << CC_HEIGHT << 24;
-        fs << CC_WIDTH << 24;
+        fs << "stageCount" << stages.size();
-        CascadeBoostParams stageParams(CvBoost::GINI, 0.999, 0.5, 0.95, 1, 200);
-        fs << CC_STAGE_PARAMS << "{"; stageParams.write( fs ); fs << "}";
-
-        fs << CC_FEATURE_PARAMS << "{";
-        fs << CC_MAX_CAT_COUNT << stages.first()->maxCatCount();
-        fs << CC_FEATURE_SIZE << 1;
-        fs << "}";
-
-        fs << CC_STAGE_NUM << stages.size();
-
-        fs << CC_STAGES << "[";
+        fs << "stages" << "[";
         foreach (const Classifier *stage, stages) {
             fs << "{";
             stage->write(fs);
+#include <opencv2/imgproc/imgproc.hpp>
+
+#include <openbr/plugins/openbr_internal.h>
+#include <openbr/core/opencvutils.h>
+
+using namespace cv;
+
+namespace br
+{
+
+#define CV_SUM_OFFSETS( p0, p1, p2, p3, rect, step )                      \
+    /* (x, y) */                                                          \
+    (p0) = (rect).x + (step) * (rect).y;                                  \
+    /* (x + w, y) */                                                      \
+    (p1) = (rect).x + (rect).width + (step) * (rect).y;                   \
+    /* (x + w, y) */                                                      \
+    (p2) = (rect).x + (step) * ((rect).y + (rect).height);                \
+    /* (x + w, y + h) */                                                  \
+    (p3) = (rect).x + (rect).width + (step) * ((rect).y + (rect).height);
+
+class HaarRepresentation : public Representation
+{
+    Q_OBJECT
+
+    Q_PROPERTY(int winWidth READ get_winWidth WRITE set_winWidth RESET reset_winWidth STORED false)
+    Q_PROPERTY(int winHeight READ get_winHeight WRITE set_winHeight RESET reset_winHeight STORED false)
+    BR_PROPERTY(int, winWidth, 24)
+    BR_PROPERTY(int, winHeight, 24)
+
+    void init()
+    {
+        int offset = winWidth + 1;
+        for (int x = 0; x < winWidth; x++) {
+            for (int y = 0; y < winHeight; y++) {
+                for (int dx = 1; dx <= winWidth; dx++) {
+                    for (int dy = 1; dy <= winHeight; dy++) {
+                        // haar_x2
+                        if ((x+dx*2 <= winWidth) && (y+dy <= winHeight))
+                            features.append(Feature(offset,
+                                                    x,    y, dx*2, dy, -1,
+                                                    x+dx, y, dx  , dy, +2));
+                        // haar_y2
+                        if ((x+dx <= winWidth) && (y+dy*2 <= winHeight))
+                            features.append(Feature(offset,
+                                                    x,    y, dx, dy*2, -1,
+                                                    x, y+dy, dx, dy,   +2));
+                        // haar_x3
+                        if ((x+dx*3 <= winWidth) && (y+dy <= winHeight))
+                            features.append(Feature(offset,
+                                                    x,    y, dx*3, dy, -1,
+                                                    x+dx, y, dx  , dy, +3));
+                        // haar_y3
+                        if ((x+dx <= winWidth) && (y+dy*3 <= winHeight))
+                            features.append(Feature(offset,
+                                                    x, y,    dx, dy*3, -1,
+                                                    x, y+dy, dx, dy,   +3));
+                        // x2_y2
+                        if ((x+dx*2 <= winWidth) && (y+dy*2 <= winHeight))
+                            features.append(Feature(offset,
+                                                    x,    y,    dx*2, dy*2, -1,
+                                                    x,    y,    dx,   dy,   +2,
+                                                    x+dx, y+dy, dx,   dy,   +2));
+
+
+                    }
+                }
+            }
+        }
+    }
+
+    void preprocess(const Mat &src, Mat &dst) const
+    {
+        integral(src, dst);
+    }
+
+    float evaluate(const Mat &image, int idx) const
+    {
+        return (float)features[idx].calc(image);
+    }
+
+    Mat evaluate(const Mat &image, 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]);
+        return result;
+    }
+
+    int numFeatures() const { return features.size(); }
+    Size preWindowSize() const { return Size(winWidth, winHeight); }
+    Size postWindowSize() const { return Size(winWidth + 1, winHeight + 1); }
+    int maxCatCount() const { return 0; }
+
+    struct Feature
+    {
+        Feature();
+        Feature( int offset,
+            int x0, int y0, int w0, int h0, float wt0,
+            int x1, int y1, int w1, int h1, float wt1,
+            int x2 = 0, int y2 = 0, int w2 = 0, int h2 = 0, float wt2 = 0.0F );
+        float calc(const Mat &img) const;
+
+        struct {
+            Rect r;
+            float weight;
+        } rect[3];
+
+        struct {
+            int p0, p1, p2, p3;
+        } fastRect[3];
+    };
+
+    QList<Feature> features;
+};
+
+BR_REGISTER(Representation, HaarRepresentation)
+
+HaarRepresentation::Feature::Feature()
+{
+    rect[0].r = rect[1].r = rect[2].r = Rect(0,0,0,0);
+    rect[0].weight = rect[1].weight = rect[2].weight = 0;
+}
+
+HaarRepresentation::Feature::Feature(int offset,
+            int x0, int y0, int w0, int h0, float wt0,
+            int x1, int y1, int w1, int h1, float wt1,
+            int x2, int y2, int w2, int h2, float wt2)
+{
+    rect[0].r.x = x0;
+    rect[0].r.y = y0;
+    rect[0].r.width  = w0;
+    rect[0].r.height = h0;
+    rect[0].weight   = wt0;
+
+    rect[1].r.x = x1;
+    rect[1].r.y = y1;
+    rect[1].r.width  = w1;
+    rect[1].r.height = h1;
+    rect[1].weight   = wt1;
+
+    rect[2].r.x = x2;
+    rect[2].r.y = y2;
+    rect[2].r.width  = w2;
+    rect[2].r.height = h2;
+    rect[2].weight   = wt2;
+
+    for (int j = 0; j < 3; j++) {
+        if( rect[j].weight == 0.0F )
+            break;
+        CV_SUM_OFFSETS(fastRect[j].p0, fastRect[j].p1, fastRect[j].p2, fastRect[j].p3, rect[j].r, offset)
+    }
+}
+
+inline float HaarRepresentation::Feature::calc(const Mat &img) const
+{
+    const int* ptr = img.ptr<int>();
+    float ret = rect[0].weight * (ptr[fastRect[0].p0] - ptr[fastRect[0].p1] - ptr[fastRect[0].p2] + ptr[fastRect[0].p3]) +
+                rect[1].weight * (ptr[fastRect[1].p0] - ptr[fastRect[1].p1] - ptr[fastRect[1].p2] + ptr[fastRect[1].p3]);
+    if (rect[2].weight != 0.0f)
+        ret += rect[2].weight * (ptr[fastRect[2].p0] - ptr[fastRect[2].p1] - ptr[fastRect[2].p2] + ptr[fastRect[2].p3]);
+    return ret;
+}
+
+} // namespace br
+
+#include "representation/haar.moc"
+
@@ -58,7 +58,6 @@ class MBLBPRepresentation : public Representation
         return result;
     }
-    void write(FileStorage &fs, const Mat &featureMap);
     int numFeatures() const { return features.size(); }
     Size preWindowSize() const { return Size(winWidth, winHeight); }
     Size postWindowSize() const { return Size(winWidth + 1, winHeight + 1); }
@@ -79,20 +78,6 @@ class MBLBPRepresentation : public Representation
 BR_REGISTER(Representation, MBLBPRepresentation)
-void MBLBPRepresentation::write(FileStorage &fs, const Mat &featureMap)
-{
-    fs << "features" << "[";
-    const Mat_<int>& featureMap_ = (const Mat_<int>&)featureMap;
-    for ( int fi = 0; fi < featureMap.cols; fi++ )
-        if ( featureMap_(0, fi) >= 0 )
-        {
-            fs << "{";
-            features[fi].write( fs );
-            fs << "}";
-        }
-    fs << "]";
-}
-
 MBLBPRepresentation::Feature::Feature( int offset, int x, int y, int _blockWidth, int _blockHeight )
 {
     Rect tr = rect = cvRect(x, y, _blockWidth, _blockHeight);