From 74f98da211d2de6f442304b3829e2458282c286c Mon Sep 17 00:00:00 2001 From: AlexeyAB Date: Sun, 4 Mar 2018 17:59:23 +0300 Subject: [PATCH] Show IoU, save anchors to file.Show anchors in the window if used -show --- build/darknet/x64/calc_anchors.cmd | 4 ++ src/detector.c | 103 ++++++++++++++++++++++++++--- 2 files changed, 97 insertions(+), 10 deletions(-) diff --git a/build/darknet/x64/calc_anchors.cmd b/build/darknet/x64/calc_anchors.cmd index d97afdd1..f8a77ad5 100644 --- a/build/darknet/x64/calc_anchors.cmd +++ b/build/darknet/x64/calc_anchors.cmd @@ -4,5 +4,9 @@ rem # How to calculate Yolo v2 anchors using K-means++ darknet.exe detector calc_anchors data/voc.data -num_of_clusters 5 -final_width 13 -final_heigh 13 +rem darknet.exe detector calc_anchors data/voc.data -num_of_clusters 5 -final_width 13 -final_heigh 13 -show + + + pause diff --git a/src/detector.c b/src/detector.c index 8eaffce8..b773a1c8 100644 --- a/src/detector.c +++ b/src/detector.c @@ -807,7 +807,7 @@ void validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float } #ifdef OPENCV -void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final_height) +void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final_height, int show) { printf("\n num_of_clusters = %d, final_width = %d, final_height = %d \n", num_of_clusters, final_width, final_height); @@ -846,7 +846,6 @@ void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final } printf("\n all loaded. \n"); - //int number_of_boxes = 10; CvMat* points = cvCreateMat(number_of_boxes, 2, CV_32FC1); CvMat* centers = cvCreateMat(num_of_clusters, 2, CV_32FC1); CvMat* labels = cvCreateMat(number_of_boxes, 1, CV_32SC1); @@ -859,7 +858,7 @@ void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final } - const int attemps = 1000; + const int attemps = 10; double compactness; enum { @@ -871,18 +870,101 @@ void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final printf("\n calculating k-means++ ..."); // Should be used: distance(box, centroid) = 1 - IoU(box, centroid) cvKMeans2(points, num_of_clusters, labels, - cvTermCriteria(CV_TERMCRIT_EPS+CV_TERMCRIT_ITER, 1000, 0), attemps, - 0, KMEANS_RANDOM_CENTERS, + cvTermCriteria(CV_TERMCRIT_EPS+CV_TERMCRIT_ITER, 10000, 0), attemps, + 0, KMEANS_PP_CENTERS, centers, &compactness); + + //orig 2.0 anchors = 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52 + //float orig_anch[] = { 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52 }; + // worse than ours (even for 19x19 final size - for input size 608x608) - printf("\n"); + //orig anchors = 1.3221,1.73145, 3.19275,4.00944, 5.05587,8.09892, 9.47112,4.84053, 11.2364,10.0071 + //float orig_anch[] = { 1.3221,1.73145, 3.19275,4.00944, 5.05587,8.09892, 9.47112,4.84053, 11.2364,10.0071 }; + // orig (IoU=59.90%) better than ours (59.75%) + + //gen_anchors.py = 1.19, 1.99, 2.79, 4.60, 4.53, 8.92, 8.06, 5.29, 10.32, 10.66 + //float orig_anch[] = { 1.19, 1.99, 2.79, 4.60, 4.53, 8.92, 8.06, 5.29, 10.32, 10.66 }; + + // ours: anchors = 9.3813,6.0095, 3.3999,5.3505, 10.9476,11.1992, 5.0161,9.8314, 1.5003,2.1595 + //float orig_anch[] = { 9.3813,6.0095, 3.3999,5.3505, 10.9476,11.1992, 5.0161,9.8314, 1.5003,2.1595 }; + //for (i = 0; i < num_of_clusters * 2; ++i) centers->data.fl[i] = orig_anch[i]; + + //for (i = 0; i < number_of_boxes; ++i) + // printf("%2.2f,%2.2f, ", points->data.fl[i * 2], points->data.fl[i * 2 + 1]); + + float avg_iou = 0; + for (i = 0; i < number_of_boxes; ++i) { + float box_w = points->data.fl[i * 2]; + float box_h = points->data.fl[i * 2 + 1]; + //int cluster_idx = labels->data.i[i]; + int cluster_idx = 0; + float min_dist = 1000000; + for (j = 0; j < num_of_clusters; ++j) { + float anchor_w = centers->data.fl[j * 2]; + float anchor_h = centers->data.fl[j * 2 + 1]; + float w_diff = anchor_w - box_w; + float h_diff = anchor_h - box_h; + float distance = sqrt(w_diff*w_diff + h_diff*h_diff); + if (distance < min_dist) min_dist = distance, cluster_idx = j; + } + + float anchor_w = centers->data.fl[cluster_idx * 2]; + float anchor_h = centers->data.fl[cluster_idx * 2 + 1]; + float min_w = (box_w < anchor_w) ? box_w : anchor_w; + float min_h = (box_h < anchor_h) ? box_h : anchor_h; + float box_intersect = min_w*min_h; + float box_union = box_w*box_h + anchor_w*anchor_h - box_intersect; + float iou = box_intersect / box_union; + if (iou > 1 || iou < 0) { + printf(" i = %d, box_w = %d, box_h = %d, anchor_w = %d, anchor_h = %d, iou = %f \n", + i, box_w, box_h, anchor_w, anchor_h, iou); + } + else avg_iou += iou; + } + avg_iou = 100 * avg_iou / number_of_boxes; + printf("\n avg IoU = %2.2f %% \n", avg_iou); + + char buff[1024]; + FILE* fw = fopen("anchors.txt", "wb"); + printf("\nSaving anchors to the file: anchors.txt \n"); printf("anchors = "); for (i = 0; i < num_of_clusters; ++i) { - printf("%2.2f,%2.2f, ", centers->data.fl[i * 2], centers->data.fl[i * 2 + 1]); + sprintf(buff, "%2.4f,%2.4f", centers->data.fl[i * 2], centers->data.fl[i * 2 + 1]); + printf("%s, ", buff); + fwrite(buff, sizeof(char), strlen(buff), fw); + if (i + 1 < num_of_clusters) fwrite(", ", sizeof(char), 2, fw);; } + printf("\n"); + fclose(fw); + + if (show) { + size_t img_size = 700; + IplImage* img = cvCreateImage(cvSize(img_size, img_size), 8, 3); + cvZero(img); + for (j = 0; j < num_of_clusters; ++j) { + CvPoint pt1, pt2; + pt1.x = pt1.y = 0; + pt2.x = centers->data.fl[j * 2] * img_size / final_width; + pt2.y = centers->data.fl[j * 2 + 1] * img_size / final_height; + cvRectangle(img, pt1, pt2, CV_RGB(255, 255, 255), 1, 8, 0); + } - //for (i = 0; i < number_of_boxes; ++i) - // printf("%2.2f,%2.2f, ", points->data.fl[i * 2], points->data.fl[i * 2 + 1]); + for (i = 0; i < number_of_boxes; ++i) { + CvPoint pt; + pt.x = points->data.fl[i * 2] * img_size / final_width; + pt.y = points->data.fl[i * 2 + 1] * img_size / final_height; + int cluster_idx = labels->data.i[i]; + int red_id = (cluster_idx * (uint64_t)123 + 55) % 255; + int green_id = (cluster_idx * (uint64_t)321 + 33) % 255; + int blue_id = (cluster_idx * (uint64_t)11 + 99) % 255; + cvCircle(img, pt, 1, CV_RGB(red_id, green_id, blue_id), CV_FILLED, 8, 0); + //if(pt.x > img_size || pt.y > img_size) printf("\n pt.x = %d, pt.y = %d \n", pt.x, pt.y); + } + cvShowImage("clusters", img); + cvWaitKey(0); + cvReleaseImage(&img); + cvDestroyAllWindows(); + } free(rel_width_height_array); cvReleaseMat(&points); @@ -961,6 +1043,7 @@ void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filenam void run_detector(int argc, char **argv) { int dont_show = find_arg(argc, argv, "-dont_show"); + int show = find_arg(argc, argv, "-show"); int http_stream_port = find_int_arg(argc, argv, "-http_port", -1); char *out_filename = find_char_arg(argc, argv, "-out_filename", 0); char *prefix = find_char_arg(argc, argv, "-prefix", 0); @@ -1010,7 +1093,7 @@ void run_detector(int argc, char **argv) else if(0==strcmp(argv[2], "valid")) validate_detector(datacfg, cfg, weights); else if(0==strcmp(argv[2], "recall")) validate_detector_recall(datacfg, cfg, weights); else if(0==strcmp(argv[2], "map")) validate_detector_map(datacfg, cfg, weights, thresh); - else if(0==strcmp(argv[2], "calc_anchors")) calc_anchors(datacfg, num_of_clusters, final_width, final_heigh); + else if(0==strcmp(argv[2], "calc_anchors")) calc_anchors(datacfg, num_of_clusters, final_width, final_heigh, show); else if(0==strcmp(argv[2], "demo")) { list *options = read_data_cfg(datacfg); int classes = option_find_int(options, "classes", 20);