diff --git a/apps/sft/include/sft/config.hpp b/apps/sft/include/sft/config.hpp index 4a1bca09ee..6e512bec45 100644 --- a/apps/sft/include/sft/config.hpp +++ b/apps/sft/include/sft/config.hpp @@ -57,6 +57,22 @@ struct Config void read(const cv::FileNode& node); + // Scaled and shrunk model size. + cv::Size model(ivector::const_iterator it) const + { + float octave = powf(2, *it); + return cv::Size( cvRound(modelWinSize.width * octave) / shrinkage, + cvRound(modelWinSize.height * octave) / shrinkage ); + } + + // Scaled but, not shrunk bounding box for object in sample image. + cv::Rect bbox(ivector::const_iterator it) const + { + float octave = powf(2, *it); + return cv::Rect( cvRound(offset.x * octave), cvRound(offset.y * octave), + cvRound(modelWinSize.width * octave), cvRound(modelWinSize.height * octave)); + } + // Paths to a rescaled data string trainPath; string testPath; diff --git a/apps/sft/include/sft/octave.hpp b/apps/sft/include/sft/octave.hpp index bddb419809..e03c29e387 100644 --- a/apps/sft/include/sft/octave.hpp +++ b/apps/sft/include/sft/octave.hpp @@ -76,12 +76,14 @@ struct ICF float operator() (const Mat& integrals, const cv::Size& model) const { - const int* ptr = integrals.ptr(0) + (model.height * channel + bb.y) * model.width + bb.x; + int step = model.width + 1; + + const int* ptr = integrals.ptr(0) + (model.height * channel + bb.y) * step + bb.x; int a = ptr[0]; int b = ptr[bb.width]; - ptr += bb.height * model.width; + ptr += bb.height * step; int c = ptr[bb.width]; int d = ptr[0]; @@ -92,13 +94,17 @@ struct ICF private: cv::Rect bb; int channel; + + friend std::ostream& operator<<(std::ostream& out, const ICF& m); }; +std::ostream& operator<<(std::ostream& out, const ICF& m); + class FeaturePool { public: FeaturePool(cv::Size model, int nfeatures); - ~FeaturePool(); + int size() const { return (int)pool.size(); } float apply(int fi, int si, const Mat& integrals) const; @@ -122,7 +128,7 @@ public: Octave(cv::Rect boundingBox, int npositives, int nnegatives, int logScale, int shrinkage); virtual ~Octave(); - virtual bool train(const Dataset& dataset, const FeaturePool& pool); + virtual bool train(const Dataset& dataset, const FeaturePool& pool, int weaks, int treeDepth); int logScale; @@ -144,7 +150,6 @@ private: Mat responses; CvBoostParams params; - }; } diff --git a/apps/sft/octave.cpp b/apps/sft/octave.cpp index 7ccd49051f..e90504f2c9 100644 --- a/apps/sft/octave.cpp +++ b/apps/sft/octave.cpp @@ -43,16 +43,6 @@ #include #include -#if defined VISUALIZE_GENERATION -# define show(a, b) \ - do { \ - cv::imshow(a,b); \ - cv::waitkey(0); \ - } while(0) -#else -# define show(a, b) -#endif - #include #include #include @@ -63,13 +53,7 @@ sft::Octave::Octave(cv::Rect bb, int np, int nn, int ls, int shr) { int maxSample = npositives + nnegatives; responses.create(maxSample, 1, CV_32FC1); -} -sft::Octave::~Octave(){} - -bool sft::Octave::train( const cv::Mat& trainData, const cv::Mat& _responses, const cv::Mat& varIdx, - const cv::Mat& sampleIdx, const cv::Mat& varType, const cv::Mat& missingDataMask) -{ CvBoostParams _params; { // tree params @@ -79,27 +63,35 @@ bool sft::Octave::train( const cv::Mat& trainData, const cv::Mat& _responses, co _params.truncate_pruned_tree = false; _params.use_surrogates = false; _params.use_1se_rule = false; - _params.regression_accuracy = 0.0; + _params.regression_accuracy = 1.0e-6; // boost params _params.boost_type = CvBoost::GENTLE; _params.split_criteria = CvBoost::SQERR; _params.weight_trim_rate = 0.95; - - /// ToDo: move to params + // simple defaults _params.min_sample_count = 2; _params.weak_count = 1; } - std::cout << "WARNING: " << sampleIdx << std::endl; - std::cout << "WARNING: " << trainData << std::endl; - std::cout << "WARNING: " << _responses << std::endl; - std::cout << "WARNING: " << varIdx << std::endl; - std::cout << "WARNING: " << varType << std::endl; + params = _params; +} + +sft::Octave::~Octave(){} + +bool sft::Octave::train( const cv::Mat& trainData, const cv::Mat& _responses, const cv::Mat& varIdx, + const cv::Mat& sampleIdx, const cv::Mat& varType, const cv::Mat& missingDataMask) +{ + + std::cout << "WARNING: sampleIdx " << sampleIdx << std::endl; + std::cout << "WARNING: trainData " << trainData << std::endl; + std::cout << "WARNING: _responses " << _responses << std::endl; + std::cout << "WARNING: varIdx" << varIdx << std::endl; + std::cout << "WARNING: varType" << varType << std::endl; bool update = false; - return cv::Boost::train(trainData, CV_COL_SAMPLE, _responses, varIdx, sampleIdx, varType, missingDataMask, _params, + return cv::Boost::train(trainData, CV_COL_SAMPLE, _responses, varIdx, sampleIdx, varType, missingDataMask, params, update); } @@ -164,29 +156,30 @@ public: }; } -// ToDo: parallelize it +// ToDo: parallelize it, fix curring // ToDo: sunch model size and shrinced model size usage/ Now model size mean already shrinked model void sft::Octave::processPositives(const Dataset& dataset, const FeaturePool& pool) { Preprocessor prepocessor(shrinkage); - int w = 64 * pow(2, logScale) /shrinkage; - int h = 128 * pow(2, logScale) /shrinkage * 10; + int w = boundingBox.width; + int h = boundingBox.height; - integrals.create(pool.size(), (w + 1) * (h + 1), CV_32SC1); + integrals.create(pool.size(), (w / shrinkage + 1) * (h / shrinkage * 10 + 1), CV_32SC1); int total = 0; - for (svector::const_iterator it = dataset.pos.begin(); it != dataset.pos.end(); ++it) { const string& curr = *it; dprintf("Process candidate positive image %s\n", curr.c_str()); - cv::Mat sample = cv::imread(curr); - cv::Mat channels = integrals.row(total).reshape(0, h + 1); - prepocessor.apply(sample, channels); + cv::Mat sample = cv::imread(curr); + cv::Mat channels = integrals.row(total).reshape(0, h / shrinkage * 10 + 1); + sample = sample(boundingBox); + + prepocessor.apply(sample, channels); responses.ptr(total)[0] = 1.f; if (++total >= npositives) break; @@ -204,8 +197,8 @@ void sft::Octave::generateNegatives(const Dataset& dataset) sft::Random::engine eng; sft::Random::engine idxEng; - int w = 64 * pow(2, logScale) /shrinkage; - int h = 128 * pow(2, logScale) /shrinkage * 10; + int w = boundingBox.width; + int h = boundingBox.height; Preprocessor prepocessor(shrinkage); @@ -222,15 +215,9 @@ void sft::Octave::generateNegatives(const Dataset& dataset) dprintf("Process %s\n", dataset.neg[curr].c_str()); Mat frame = cv::imread(dataset.neg[curr]); - prepocessor.apply(frame, sum); - std::cout << "WARNING: " << frame.cols << " " << frame.rows << std::endl; - std::cout << "WARNING: " << frame.cols / shrinkage << " " << frame.rows / shrinkage << std::endl; - - int maxW = frame.cols / shrinkage - 2 * boundingBox.x - boundingBox.width; - int maxH = frame.rows / shrinkage - 2 * boundingBox.y - boundingBox.height; - - std::cout << "WARNING: " << maxW << " " << maxH << std::endl; + int maxW = frame.cols - 2 * boundingBox.x - boundingBox.width; + int maxH = frame.rows - 2 * boundingBox.y - boundingBox.height; sft::Random::uniform wRand(0, maxW -1); sft::Random::uniform hRand(0, maxH -1); @@ -238,19 +225,16 @@ void sft::Octave::generateNegatives(const Dataset& dataset) int dx = wRand(eng); int dy = hRand(eng); - std::cout << "WARNING: " << dx << " " << dy << std::endl; - std::cout << "WARNING: " << dx + boundingBox.width + 1 << " " << dy + boundingBox.height + 1 << std::endl; - std::cout << "WARNING: " << sum.cols << " " << sum.rows << std::endl; + frame = frame(cv::Rect(dx, dy, boundingBox.width, boundingBox.height)); - sum = sum(cv::Rect(dx, dy, boundingBox.width + 1, boundingBox.height * 10 + 1)); + cv::Mat channels = integrals.row(i).reshape(0, h / shrinkage * 10 + 1); + prepocessor.apply(frame, channels); dprintf("generated %d %d\n", dx, dy); - // if (predict(sum)) + // // if (predict(sum)) { responses.ptr(i)[0] = 0.f; - // sum = sum.reshape(0, 1); - sum.copyTo(integrals.row(i).reshape(0, h + 1)); ++i; } } @@ -258,11 +242,18 @@ void sft::Octave::generateNegatives(const Dataset& dataset) dprintf("Processing negatives finished:\n\trequested %d negatives, viewed %d samples.\n", nnegatives, total); } -bool sft::Octave::train(const Dataset& dataset, const FeaturePool& pool) +bool sft::Octave::train(const Dataset& dataset, const FeaturePool& pool, int weaks, int treeDepth) { + CV_Assert(treeDepth == 2); + CV_Assert(weaks > 0); + + params.max_depth = treeDepth; + params.weak_count = weaks; + // 1. fill integrals and classes processPositives(dataset, pool); generateNegatives(dataset); + // exit(0); // 2. only sumple case (all features used) int nfeatures = pool.size(); @@ -313,8 +304,6 @@ sft::FeaturePool::FeaturePool(cv::Size m, int n) : model(m), nfeatures(n) fill(nfeatures); } -sft::FeaturePool::~FeaturePool(){} - float sft::FeaturePool::apply(int fi, int si, const Mat& integrals) const { return pool[fi](integrals.row(si), model); @@ -323,13 +312,13 @@ float sft::FeaturePool::apply(int fi, int si, const Mat& integrals) const void sft::FeaturePool::fill(int desired) { - int mw = model.width; int mh = model.height; int maxPoolSize = (mw -1) * mw / 2 * (mh - 1) * mh / 2 * N_CHANNELS; nfeatures = std::min(desired, maxPoolSize); + dprintf("Requeste feature pool %d max %d suggested %d\n", desired, maxPoolSize, nfeatures); pool.reserve(nfeatures); @@ -363,10 +352,19 @@ void sft::FeaturePool::fill(int desired) sft::ICF f(x, y, w, h, ch); if (std::find(pool.begin(), pool.end(),f) == pool.end()) + { + // std::cout << f << std::endl; pool.push_back(f); + } } } +std::ostream& sft::operator<<(std::ostream& out, const sft::ICF& m) +{ + out << m.channel << " " << m.bb; + return out; +} + // ============ Dataset ============ // namespace { using namespace sft; diff --git a/apps/sft/sft.cpp b/apps/sft/sft.cpp index 8a96b31023..6ea3513473 100644 --- a/apps/sft/sft.cpp +++ b/apps/sft/sft.cpp @@ -106,47 +106,34 @@ int main(int argc, char** argv) // 3. Train all octaves for (ivector::const_iterator it = cfg.octaves.begin(); it != cfg.octaves.end(); ++it) { + // a. create rangom feature pool int nfeatures = cfg.poolSize; + cv::Size model = cfg.model(it); + std::cout << "Model " << model << std::endl; + sft::FeaturePool pool(model, nfeatures); + nfeatures = pool.size(); + + int npositives = cfg.positives; int nnegatives = cfg.negatives; - int shrinkage = cfg.shrinkage; - int octave = *it; + cv::Rect boundingBox = cfg.bbox(it); + std::cout << "Object bounding box" << boundingBox << std::endl; + + sft::Octave boost(boundingBox, npositives, nnegatives, *it, shrinkage); - cv::Size model = cv::Size(cfg.modelWinSize.width / cfg.shrinkage, cfg.modelWinSize.height / cfg.shrinkage ); std::string path = cfg.trainPath; - - cv::Rect boundingBox(cfg.offset.x / cfg.shrinkage, cfg.offset.y / cfg.shrinkage, - cfg.modelWinSize.width / cfg.shrinkage, cfg.modelWinSize.height / cfg.shrinkage); - - sft::Octave boost(boundingBox, npositives, nnegatives, octave, shrinkage); - - sft::FeaturePool pool(model, nfeatures); sft::Dataset dataset(path, boost.logScale); - if (boost.train(dataset, pool)) + if (boost.train(dataset, pool, cfg.weaks, cfg.treeDepth)) { - } - std::cout << "Octave " << octave << " was successfully trained..." << std::endl; - // // d. crain octave - // if (octave.train(pool, cfg.positives, cfg.negatives, cfg.weaks)) - // { + std::cout << "Octave " << *it << " was successfully trained..." << std::endl; // strong.push_back(octave); - // } + } } // fso << "]" << "}"; -// // 3. create Soft Cascade -// // sft::SCascade cascade(cfg.modelWinSize, cfg.octs, cfg.shrinkage); - -// // // 4. Generate feature pool -// // std::vector pool; -// // sft::fillPool(pool, cfg.poolSize, cfg.modelWinSize / cfg.shrinkage, cfg.seed); - -// // // 5. Train all octaves -// // cascade.train(cfg.trainPath); - // // // 6. Set thresolds // // cascade.prune();