From c704942b8ae217f196e1220d293b17f1d16a9d7c Mon Sep 17 00:00:00 2001 From: Vladislav Sovrasov Date: Tue, 17 Oct 2017 11:24:50 +0300 Subject: [PATCH] dnn: add a documentation for NMS, fix missing experimantal namespace --- modules/dnn/include/opencv2/dnn/dnn.hpp | 20 +++++++++------- modules/dnn/include/opencv2/dnn/nms.inl.hpp | 24 ++++--------------- .../dnn/src/layers/detection_output_layer.cpp | 18 ++++---------- modules/dnn/src/nms.cpp | 13 ++++++++-- 4 files changed, 32 insertions(+), 43 deletions(-) diff --git a/modules/dnn/include/opencv2/dnn/dnn.hpp b/modules/dnn/include/opencv2/dnn/dnn.hpp index d276abf099..7f1a2b3630 100644 --- a/modules/dnn/include/opencv2/dnn/dnn.hpp +++ b/modules/dnn/include/opencv2/dnn/dnn.hpp @@ -734,18 +734,20 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN */ CV_EXPORTS_W void shrinkCaffeModel(const String& src, const String& dst); - /** @brief - * @param bboxes - * @param scores - * @param score_threshold - * @param nms_threshold - * @param eta - * @param top_k - * @param indices + /** @brief Performs non maximum suppression given boxes and corresponding scores. + + * @param bboxes a set of bounding boxes to apply NMS. + * @param scores a set of corresponding confidences. + * @param score_threshold a threshold used to filter boxes by score. + * @param nms_threshold a threshold used in non maximum suppression. + * @param indices the kept indices of bboxes after NMS. + * @param eta a coefficient in adaptive threshold formula: \f$nms\_threshold_{i+1}=eta\cdot nms\_threshold_i\f$. + * @param top_k if `>0`, keep at most @p top_k picked indices. */ CV_EXPORTS_W void NMSBoxes(const std::vector& bboxes, const std::vector& scores, const float score_threshold, const float nms_threshold, - const float eta, const int top_k, CV_OUT std::vector& indices); + CV_OUT std::vector& indices, + const float eta = 1.f, const int top_k = 0); //! @} diff --git a/modules/dnn/include/opencv2/dnn/nms.inl.hpp b/modules/dnn/include/opencv2/dnn/nms.inl.hpp index 26183a083a..89e3adfcf5 100644 --- a/modules/dnn/include/opencv2/dnn/nms.inl.hpp +++ b/modules/dnn/include/opencv2/dnn/nms.inl.hpp @@ -48,25 +48,12 @@ inline void GetMaxScoreIndex(const std::vector& scores, const float thres SortScorePairDescend); // Keep top_k scores if needed. - if (top_k > -1 && top_k < (int)score_index_vec.size()) + if (top_k > 0 && top_k < (int)score_index_vec.size()) { score_index_vec.resize(top_k); } } -template -struct NMSOverlap -{ - float operator() (const BoxType& a, const BoxType& b); -}; - -template <> -inline float NMSOverlap::operator() (const Rect& a, const Rect& b) -{ - float rectIntersectionArea = (float)(a & b).area(); - return rectIntersectionArea / (a.area() + b.area() - rectIntersectionArea); -} - // Do non maximum suppression given bboxes and scores. // Inspired by Piotr Dollar's NMS implementation in EdgeBox. // https://goo.gl/jV3JYS @@ -74,13 +61,13 @@ inline float NMSOverlap::operator() (const Rect& a, const Rect& b) // scores: a set of corresponding confidences. // score_threshold: a threshold used to filter detection results. // nms_threshold: a threshold used in non maximum suppression. -// top_k: if not -1, keep at most top_k picked indices. +// top_k: if not > 0, keep at most top_k picked indices. // indices: the kept indices of bboxes after nms. template inline void NMSFast_(const std::vector& bboxes, const std::vector& scores, const float score_threshold, const float nms_threshold, const float eta, const int top_k, - std::vector& indices, NMSOverlap computeOverlap) + std::vector& indices, float (*computeOverlap)(const BoxType&, const BoxType&)) { CV_Assert(bboxes.size() == scores.size()); @@ -91,8 +78,8 @@ inline void NMSFast_(const std::vector& bboxes, // Do nms. float adaptive_threshold = nms_threshold; indices.clear(); - while (score_index_vec.size() != 0) { - const int idx = score_index_vec.front().second; + for (size_t i = 0; i < score_index_vec.size(); ++i) { + const int idx = score_index_vec[i].second; bool keep = true; for (int k = 0; k < (int)indices.size() && keep; ++k) { const int kept_idx = indices[k]; @@ -101,7 +88,6 @@ inline void NMSFast_(const std::vector& bboxes, } if (keep) indices.push_back(idx); - score_index_vec.erase(score_index_vec.begin()); if (keep && eta < 1 && adaptive_threshold > 0.5) { adaptive_threshold *= eta; } diff --git a/modules/dnn/src/layers/detection_output_layer.cpp b/modules/dnn/src/layers/detection_output_layer.cpp index 1312a81b13..970984b341 100644 --- a/modules/dnn/src/layers/detection_output_layer.cpp +++ b/modules/dnn/src/layers/detection_output_layer.cpp @@ -62,6 +62,8 @@ static inline bool SortScorePairDescend(const std::pair& pair1, return pair1.first > pair2.first; } +static inline float caffe_box_overlap(const caffe::NormalizedBBox& a, const caffe::NormalizedBBox& b); + } // namespace class DetectionOutputLayerImpl : public DetectionOutputLayer @@ -309,7 +311,8 @@ public: LabelBBox::const_iterator label_bboxes = decodeBBoxes.find(label); if (label_bboxes == decodeBBoxes.end()) CV_ErrorNoReturn_(cv::Error::StsError, ("Could not find location predictions for label %d", label)); - ApplyNMSFast(label_bboxes->second, scores, _confidenceThreshold, _nmsThreshold, 1.0, _topK, indices[c]); + NMSFast_(label_bboxes->second, scores, _confidenceThreshold, _nmsThreshold, 1.0, _topK, + indices[c], util::caffe_box_overlap); numDetections += indices[c].size(); } if (_keepTopK > -1 && numDetections > (size_t)_keepTopK) @@ -620,16 +623,6 @@ public: } } - - - static void ApplyNMSFast(const std::vector& bboxes, - const std::vector& scores, const float score_threshold, - const float nms_threshold, const float eta, const int top_k, - std::vector& indices) - { - NMSFast_(bboxes, scores, score_threshold, nms_threshold, eta, top_k, indices, NMSOverlap()); - } - // Compute the jaccard (intersection over union IoU) overlap between two bboxes. template static float JaccardOverlap(const caffe::NormalizedBBox& bbox1, @@ -675,8 +668,7 @@ public: } }; -template <> -float NMSOverlap::operator() (const caffe::NormalizedBBox& a, const caffe::NormalizedBBox& b) +float util::caffe_box_overlap(const caffe::NormalizedBBox& a, const caffe::NormalizedBBox& b) { return DetectionOutputLayerImpl::JaccardOverlap(a, b); } diff --git a/modules/dnn/src/nms.cpp b/modules/dnn/src/nms.cpp index 5f433dba22..af9e9c855c 100644 --- a/modules/dnn/src/nms.cpp +++ b/modules/dnn/src/nms.cpp @@ -12,13 +12,22 @@ namespace cv { namespace dnn { +CV__DNN_EXPERIMENTAL_NS_BEGIN + +static inline float rectOverlap(const Rect& a, const Rect& b) +{ + return 1.f - static_cast(jaccardDistance(a, b)); +} void NMSBoxes(const std::vector& bboxes, const std::vector& scores, const float score_threshold, const float nms_threshold, - const float eta, const int top_k, std::vector& indices) + std::vector& indices, const float eta, const int top_k) { - NMSFast_(bboxes, scores, score_threshold, nms_threshold, eta, top_k, indices, NMSOverlap()); + CV_Assert(bboxes.size() == scores.size(), score_threshold >= 0, + nms_threshold >= 0, eta > 0); + NMSFast_(bboxes, scores, score_threshold, nms_threshold, eta, top_k, indices, rectOverlap); } +CV__DNN_EXPERIMENTAL_NS_END }// dnn }// cv