Fix some bugs

Author: Nuzhny007

example/VideoExample.cpp

@@ -151,6 +151,8 @@ void VideoExample::SyncProcess()
 
 		FrameInfo frameInfo(m_batchSize);
 		frameInfo.m_frames.emplace_back(frame);
+		frameInfo.m_frameInds.emplace_back(framesCounter);
+		++framesCounter;
 		for (size_t i = 1; i < m_batchSize; ++i)
 		{
 			capture >> frame;
@@ -366,8 +368,7 @@ void VideoExample::CaptureAndDetect(VideoExample* thisPtr, std::atomic<bool>& st
 				frameInfo.m_cond.notify_one();
 				break;
 			}
-			frameInfo.m_frameInds.emplace_back(framesCounter);
-
+			frameInfo.m_frameInds[i] = framesCounter;
 			++framesCounter;
 		}
 		if (frameInfo.m_frames.size() < frameInfo.m_batchSize)
@@ -419,8 +420,9 @@ void VideoExample::Detection(FrameInfo& frame)
         if (m_detector->CanGrayProcessing())
             frames.emplace_back(frame.m_frames[i].GetUMatGray());
         else
-            frames.emplace_back(frame.m_frames[i].GetUMatBGR());
+            frames.push_back(frame.m_frames[i].GetUMatBGR());
 	}
+	frame.CleanRegions();
     m_detector->Detect(frames, frame.m_regions);
 }
 
@@ -431,6 +433,8 @@ void VideoExample::Detection(FrameInfo& frame)
 ///
 void VideoExample::Tracking(FrameInfo& frame)
 {
+	assert(frame.m_regions.size() == frame.m_frames.size());
+
 	for (size_t i = 0; i < frame.m_frames.size(); ++i)
 	{
 		if (m_tracker->CanColorFrameToTrack())

example/VideoExample.h

@@ -197,7 +197,7 @@ class Frame
 	{
 		if (m_umBGR.empty())
 			m_umBGR = m_mBGR.getUMat(cv::ACCESS_READ);
-		return m_umGray;
+		return m_umBGR;
 	}
 	///
 	const cv::UMat& GetUMatGray()
@@ -256,6 +256,17 @@ struct FrameInfo
 		m_frameInds.reserve(m_batchSize);
 	}
 
+	///
+	void CleanRegions()
+	{
+		if (m_regions.size() != m_batchSize)
+			m_regions.resize(m_batchSize);
+		for (auto& reg : m_regions)
+		{
+			reg.clear();
+		}
+	}
+
 	std::vector<Frame> m_frames;
 	std::vector<regions_t> m_regions;
 	std::vector<int> m_frameInds;

src/Detector/YoloDarknetDetector.cpp

@@ -145,7 +145,7 @@ void YoloDarknetDetector::Detect(const cv::UMat& colorFrame)
 			}
 		}
 
-		if (crops.size() > 1)
+		if (crops.size() > 1 || m_batchSize > 1)
 		{
 			nms3<CRegion>(tmpRegions, m_regions, 0.4f,
 				[](const CRegion& reg) { return reg.m_brect; },
@@ -305,25 +305,42 @@ void YoloDarknetDetector::FillBatchImg(const std::vector<cv::Mat>& batch, image_
 ///
 void YoloDarknetDetector::Detect(const std::vector<cv::UMat>& frames, std::vector<regions_t>& regions)
 {
-	std::vector<cv::Mat> batch;
-	for (const auto& frame : frames)
+	if (frames.size() == 1)
 	{
-		batch.emplace_back(frame.getMat(cv::ACCESS_READ));
+		Detect(frames[0].getMat(cv::ACCESS_READ), regions[0]);
 	}
+	else
+	{
+		std::vector<cv::Mat> batch;
+		for (const auto& frame : frames)
+		{
+			batch.emplace_back(frame.getMat(cv::ACCESS_READ));
+		}
 
-	image_t detImage;
-	FillBatchImg(batch, detImage);
-	std::vector<std::vector<bbox_t>> result_vec = m_detector->detectBatch(detImage, static_cast<int>(frames.size()), m_netSize.width, m_netSize.height, m_confidenceThreshold);
+		image_t detImage;
+		FillBatchImg(batch, detImage);
+		std::vector<std::vector<bbox_t>> result_vec = m_detector->detectBatch(detImage, static_cast<int>(frames.size()), m_netSize.width, m_netSize.height, m_confidenceThreshold);
 
-	float wk = static_cast<float>(frames[0].cols) / m_netSize.width;
-	float hk = static_cast<float>(frames[0].rows) / m_netSize.height;
-	for (size_t i = 0; i < regions.size(); ++i)
-	{
-		for (const auto& bbox : result_vec[i])
+		regions_t tmpRegions;
+		tmpRegions.reserve(result_vec[0].size() + 16);
+		float wk = static_cast<float>(frames[0].cols) / m_netSize.width;
+		float hk = static_cast<float>(frames[0].rows) / m_netSize.height;
+		for (size_t i = 0; i < regions.size(); ++i)
 		{
-			if (m_classesWhiteList.empty() || m_classesWhiteList.find(T2T(bbox.obj_id)) != std::end(m_classesWhiteList))
-				regions[i].emplace_back(cv::Rect(cvRound(wk * bbox.x), cvRound(hk * bbox.y), cvRound(wk * bbox.w), cvRound(hk * bbox.h)), T2T(bbox.obj_id), bbox.prob);
+			tmpRegions.clear();
+			for (const auto& bbox : result_vec[i])
+			{
+				if (m_classesWhiteList.empty() || m_classesWhiteList.find(T2T(bbox.obj_id)) != std::end(m_classesWhiteList))
+					tmpRegions.emplace_back(cv::Rect(cvRound(wk * bbox.x), cvRound(hk * bbox.y), cvRound(wk * bbox.w), cvRound(hk * bbox.h)), T2T(bbox.obj_id), bbox.prob);
+			}
+
+			nms3<CRegion>(tmpRegions, regions[i], 0.4f,
+				[](const CRegion& reg) { return reg.m_brect; },
+				[](const CRegion& reg) { return reg.m_confidence; },
+				[](const CRegion& reg) { return reg.m_type; },
+				0, 0.f);
 		}
+
+		m_regions.assign(std::begin(regions.back()), std::end(regions.back()));
 	}
-	m_regions.assign(std::begin(regions.back()), std::end(regions.back()));
 }

src/Detector/darknet/src/yolo_v2_class.cpp

@@ -361,7 +361,7 @@ LIB_API std::vector<std::vector<bbox_t>> Detector::detectBatch(image_t img, int
     in_img.data = img.data;
     det_num_pair* prediction = network_predict_batch(&net, in_img, batch_size, width, height, thresh, hier_thresh, 0, 0, 0);
 
-    std::vector<std::vector<bbox_t>> bbox_vec;
+    std::vector<std::vector<bbox_t>> bbox_vec(batch_size);
 
     for (int bi = 0; bi < batch_size; ++bi)
     {