Output improvements for detector results: When printing detector results, o...

IlyaOvodov

2018-05-08 1b5afb45838e603fa6780762eb8cc59246dc2d81

Output improvements for detector results:
When printing detector results, output was done in random order, obfuscating results for interpreting. Now:
1. Text output includes coordinates of rects in (left,right,top,bottom in pixels) along with label and score
2. Text output is sorted by rect lefts to simplify finding appropriate rects on image
3. If several class probs are > thresh for some detection, the most probable is written first and coordinates for others are not repeated
4. Rects are imprinted in image in order by their best class prob, so most probable rects are always on top and not overlayed by less probable ones
5. Most probable label for rect is always written first
Also:
6. Message about low GPU memory include required amount

Upd:
* struct detection changes reverted to keep it unchanged
* cmd line key "-ext_output" is added to manage extended output (rect coords in detector test mode)

# Conflicts:
#	src/detector.c

 src/box.h

@@ -30,10 +30,14 @@
    float *mask;
    float objectness;
    int sort_class;
} detection;

typedef struct detection_with_class {
    detection det;
    // The most probable class id: the best class index in this->prob.
    // Is filled temporary when processing results, otherwise not initialized
    int best_class;
} detection;
} detection_with_class;

box float_to_box(float *f);
float box_iou(box a, box b);
@@ -46,4 +50,8 @@
box decode_box(box b, box anchor);
box encode_box(box b, box anchor);

// Creates array of detections with prob > thresh and fills best_class for them
// Return number of selected detections in *selected_detections_num
detection_with_class* get_actual_detections(detection *dets, int dets_num, float thresh, int* selected_detections_num);

#endif

 src/darknet.c

@@ -13,7 +13,7 @@
#endif

extern void predict_classifier(char *datacfg, char *cfgfile, char *weightfile, char *filename, int top);
extern void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh);
extern void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh, int ext_output);
extern void run_voxel(int argc, char **argv);
extern void run_yolo(int argc, char **argv);
extern void run_detector(int argc, char **argv);
@@ -382,8 +382,9 @@
        run_detector(argc, argv);
    } else if (0 == strcmp(argv[1], "detect")){
        float thresh = find_float_arg(argc, argv, "-thresh", .24);
        int ext_output = find_arg(argc, argv, "-ext_output");
        char *filename = (argc > 4) ? argv[4]: 0;
        test_detector("cfg/coco.data", argv[2], argv[3], filename, thresh);
        test_detector("cfg/coco.data", argv[2], argv[3], filename, thresh, ext_output);
    } else if (0 == strcmp(argv[1], "cifar")){
        run_cifar(argc, argv);
    } else if (0 == strcmp(argv[1], "go")){

 src/detector.c

@@ -1036,7 +1036,8 @@
}
#endif // OPENCV

void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh, float hier_thresh, int dont_show)
void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh,
                   float hier_thresh, int dont_show, int ext_output)
{
    list *options = read_data_cfg(datacfg);
    char *name_list = option_find_str(options, "names", "data/names.list");
@@ -1088,7 +1089,7 @@
        int nboxes = 0;
        detection *dets = get_network_boxes(&net, im.w, im.h, thresh, hier_thresh, 0, 1, &nboxes, letterbox);
        if (nms) do_nms_sort(dets, nboxes, l.classes, nms);
        draw_detections_v3(im, dets, nboxes, thresh, names, alphabet, l.classes);
        draw_detections_v3(im, dets, nboxes, thresh, names, alphabet, l.classes, ext_output);
        free_detections(dets, nboxes);
        save_image(im, "predictions");
        if (!dont_show) {
@@ -1124,6 +1125,9 @@
    int num_of_clusters = find_int_arg(argc, argv, "-num_of_clusters", 5);
    int width = find_int_arg(argc, argv, "-width", -1);
    int height = find_int_arg(argc, argv, "-height", -1);
    // extended output in test mode (output of rect bound coords)
    // and for recall mode (extended output table-like format with results for best_class fit)
    int ext_output = find_arg(argc, argv, "-ext_output");
    if(argc < 4){
        fprintf(stderr, "usage: %s %s [train/test/valid] [cfg] [weights (optional)]\n", argv[0], argv[1]);
        return;
@@ -1160,7 +1164,7 @@
        if(strlen(weights) > 0)
            if (weights[strlen(weights) - 1] == 0x0d) weights[strlen(weights) - 1] = 0;
    char *filename = (argc > 6) ? argv[6]: 0;
    if(0==strcmp(argv[2], "test")) test_detector(datacfg, cfg, weights, filename, thresh, hier_thresh, dont_show);
    if(0==strcmp(argv[2], "test")) test_detector(datacfg, cfg, weights, filename, thresh, hier_thresh, dont_show, ext_output);
    else if(0==strcmp(argv[2], "train")) train_detector(datacfg, cfg, weights, gpus, ngpus, clear, dont_show);
    else if(0==strcmp(argv[2], "valid")) validate_detector(datacfg, cfg, weights, outfile);
    else if(0==strcmp(argv[2], "recall")) validate_detector_recall(datacfg, cfg, weights);

 src/image.c

@@ -232,21 +232,22 @@


// Creates array of detections with prob > thresh and fills best_class for them
detection** get_actual_detections(detection *dets, int dets_num, int classes, float thresh, int* selected_detections_num)
detection_with_class* get_actual_detections(detection *dets, int dets_num, float thresh, int* selected_detections_num)
{
    int selected_num = 0;
    detection** result_arr = calloc(dets_num, sizeof(detection*));
    detection_with_class* result_arr = calloc(dets_num, sizeof(detection_with_class));
    for (int i = 0; i < dets_num; ++i) {
        dets[i].best_class = -1;
        for (int j = 0; j < classes; ++j) {
            if (dets[i].prob[j] > thresh) {
                if (dets[i].best_class < 0 || dets[i].prob[dets[i].best_class] < dets[i].prob[j]) {
                    dets[i].best_class = j;
        int best_class = -1;
        float best_class_prob = thresh;
        for (int j = 0; j < dets[i].classes; ++j) {
            if (dets[i].prob[j] > best_class_prob ) {
                best_class = j;
                best_class_prob = dets[i].prob[j];
                }
            }
        }
        if (dets[i].best_class >= 0) {
            result_arr[selected_num] = &(dets[i]);
        if (best_class >= 0) {
            result_arr[selected_num].det = dets[i];
            result_arr[selected_num].best_class = best_class;
            ++selected_num;
        }
    }
@@ -256,36 +257,41 @@
}

// compare to sort detection** by bbox.x
int compare_by_lefts(const void *a, const void *b) {
    const float delta = ((*(detection**)a)->bbox.x - (*(detection**)a)->bbox.w / 2) - ((*(detection**)b)->bbox.x - (*(detection**)b)->bbox.w / 2);
int compare_by_lefts(const void *a_ptr, const void *b_ptr) {
    const detection_with_class* a = (detection_with_class*)a_ptr;
    const detection_with_class* b = (detection_with_class*)b_ptr;
    const float delta = (a->det.bbox.x - a->det.bbox.w/2) - (b->det.bbox.x - b->det.bbox.w/2);
    return delta < 0 ? -1 : delta > 0 ? 1 : 0;
}

// compare to sort detection** by best_class probability 
int compare_by_probs(const void *a_ptr, const void *b_ptr) {
    const detection* a = *(detection**)a_ptr;
    const detection* b = *(detection**)b_ptr;
    float delta = a->prob[a->best_class] - b->prob[b->best_class];
    const detection_with_class* a = (detection_with_class*)a_ptr;
    const detection_with_class* b = (detection_with_class*)b_ptr;
    float delta = a->det.prob[a->best_class] - b->det.prob[b->best_class];
    return delta < 0 ? -1 : delta > 0 ? 1 : 0;
}

void draw_detections_v3(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes)
void draw_detections_v3(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes, int ext_output)
{
    int selected_detections_num;
    detection** selected_detections = get_actual_detections(dets, num, classes, thresh, &selected_detections_num);
    detection_with_class* selected_detections = get_actual_detections(dets, num, thresh, &selected_detections_num);

    // text output
    qsort(selected_detections, selected_detections_num, sizeof(*selected_detections), compare_by_lefts);
    for (int i = 0; i < selected_detections_num; ++i) {
        const int best_class = selected_detections[i]->best_class;
        printf("%s: %.0f%%\t(left: %.0f\ttop: %.0f\tw: %0.f\th: %0.f)\n", names[best_class],
            selected_detections[i]->prob[best_class] * 100,
            (selected_detections[i]->bbox.x - selected_detections[i]->bbox.w / 2)*im.w,
            (selected_detections[i]->bbox.y - selected_detections[i]->bbox.h / 2)*im.h,
            selected_detections[i]->bbox.w*im.w, selected_detections[i]->bbox.h*im.h);
        const int best_class = selected_detections[i].best_class;
        printf("%s: %.0f%%", names[best_class], selected_detections[i].det.prob[best_class] * 100);
        if (ext_output)
            printf("\t(left: %.0f\ttop: %.0f\tw: %0.f\th: %0.f)\n",
                (selected_detections[i].det.bbox.x - selected_detections[i].det.bbox.w / 2)*im.w,
                (selected_detections[i].det.bbox.y - selected_detections[i].det.bbox.h / 2)*im.h,
                selected_detections[i].det.bbox.w*im.w, selected_detections[i].det.bbox.h*im.h);
        else
            printf("\n");
        for (int j = 0; j < classes; ++j) {
            if (selected_detections[i]->prob[j] > thresh && j != best_class) {
                printf("%s: %.0f%%\n", names[j], selected_detections[i]->prob[j] * 100);
            if (selected_detections[i].det.prob[j] > thresh && j != best_class) {
                printf("%s: %.0f%%\n", names[j], selected_detections[i].det.prob[j] * 100);
            }
        }
    }
@@ -304,8 +310,8 @@
            }
            */

            //printf("%d %s: %.0f%%\n", i, names[dets[i].best_class], prob*100);
            int offset = selected_detections[i]->best_class * 123457 % classes;
            //printf("%d %s: %.0f%%\n", i, names[selected_detections[i].best_class], prob*100);
            int offset = selected_detections[i].best_class * 123457 % classes;
            float red = get_color(2, offset, classes);
            float green = get_color(1, offset, classes);
            float blue = get_color(0, offset, classes);
@@ -316,7 +322,7 @@
            rgb[0] = red;
            rgb[1] = green;
            rgb[2] = blue;
            box b = selected_detections[i]->bbox;
            box b = selected_detections[i].det.bbox;
            //printf("%f %f %f %f\n", b.x, b.y, b.w, b.h);

            int left = (b.x - b.w / 2.)*im.w;
@@ -338,9 +344,9 @@
            draw_box_width(im, left, top, right, bot, width, red, green, blue);
            if (alphabet) {
                char labelstr[4096] = { 0 };
                strcat(labelstr, names[selected_detections[i]->best_class]);
                strcat(labelstr, names[selected_detections[i].best_class]);
                for (int j = 0; j < classes; ++j) {
                    if (selected_detections[i]->prob[j] > thresh && j != selected_detections[i]->best_class) {
                    if (selected_detections[i].det.prob[j] > thresh && j != selected_detections[i].best_class) {
                        strcat(labelstr, ", ");
                        strcat(labelstr, names[j]);
                    }
@@ -349,8 +355,8 @@
                draw_label(im, top + width, left, label, rgb);
                free_image(label);
            }
            if (selected_detections[i]->mask) {
                image mask = float_to_image(14, 14, 1, selected_detections[i]->mask);
            if (selected_detections[i].det.mask) {
                image mask = float_to_image(14, 14, 1, selected_detections[i].det.mask);
                image resized_mask = resize_image(mask, b.w*im.w, b.h*im.h);
                image tmask = threshold_image(resized_mask, .5);
                embed_image(tmask, im, left, top);

 src/image.h

@@ -23,7 +23,7 @@
void draw_label(image a, int r, int c, image label, const float *rgb);
void write_label(image a, int r, int c, image *characters, char *string, float *rgb);
void draw_detections(image im, int num, float thresh, box *boxes, float **probs, char **names, image **labels, int classes);
void draw_detections_v3(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes);
void draw_detections_v3(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes, int ext_output);
image image_distance(image a, image b);
void scale_image(image m, float s);
image crop_image(image im, int dx, int dy, int w, int h);