Single image feature extraction for VOC

Joseph Redmon

2014-02-19 2c6d4ba1d5cffd26c5c9527175d565a81226e18d

 src/tests.c

@@ -366,20 +366,21 @@

void train_VOC()
{
    network net = parse_network_cfg("cfg/voc_backup_ramp_80.cfg");
    network net = parse_network_cfg("cfg/voc_start.cfg");
    srand(2222222);
    int i = 0;
    int i = 20;
    char *labels[] = {"aeroplane","bicycle","bird","boat","bottle","bus","car","cat","chair","cow","diningtable","dog","horse","motorbike","person","pottedplant","sheep","sofa","train","tvmonitor"};
    float lr = .00001;
    float momentum = .9;
    float decay = 0.01;
    while(i++ < 1000 || 1){
        visualize_network(net);
        cvWaitKey(100);
        data train = load_data_image_pathfile_random("images/VOC2012/train_paths.txt", 1000, labels, 20, 300, 400);
        data train = load_data_image_pathfile_random("images/VOC2012/val_paths.txt", 1000, labels, 20, 300, 400);

        image im = float_to_image(300, 400, 3,train.X.vals[0]);
        show_image(im, "input");
        visualize_network(net);
        cvWaitKey(100);

        normalize_data_rows(train);
        clock_t start = clock(), end;
        float loss = train_network_sgd(net, train, 1000, lr, momentum, decay);
@@ -388,14 +389,175 @@
        free_data(train);
        if(i%10==0){
            char buff[256];
            sprintf(buff, "cfg/voc_backup_ramp_%d.cfg", i);
            sprintf(buff, "cfg/voc_clean_ramp_%d.cfg", i);
            save_network(net, buff);
        }
        //lr *= .99;
    }
}

int main()
int voc_size(int x)
{
    x = x-1+3;
    x = x-1+3;
    x = (x-1)*2+1;
    x = x-1+5;
    x = (x-1)*2+1;
    x = (x-1)*4+11;
    return x;
}

image features_output_size(network net, IplImage *src, int outh, int outw)
{
    int h = voc_size(outh);
    int w = voc_size(outw);

    IplImage *sized = cvCreateImage(cvSize(w,h), src->depth, src->nChannels);
    cvResize(src, sized, CV_INTER_LINEAR);
    image im = ipl_to_image(sized);
    reset_network_size(net, im.h, im.w, im.c);
    forward_network(net, im.data);
    image out = get_network_image_layer(net, 5);
    //printf("%d %d\n%d %d\n", outh, out.h, outw, out.w);
    free_image(im);
    cvReleaseImage(&sized);
    return copy_image(out);
}

void features_VOC(int part, int total)
{
    int i,j, count = 0;
    network net = parse_network_cfg("cfg/voc_features.cfg");
    char *path_file = "images/VOC2012/all_paths.txt";
    char *out_dir = "voc_features/";
    list *paths = get_paths(path_file);
    node *n = paths->front;
    int size = paths->size;
    for(count = 0; count < part*size/total; ++count) n = n->next;
    while(n && count++ < (part+1)*size/total){
        char *path = (char *)n->val;
        char buff[1024];
        sprintf(buff, "%s%s.txt",out_dir, path);
        printf("%s\n", path);
        FILE *fp = fopen(buff, "w");
        if(fp == 0) file_error(buff);

        IplImage* src = 0;
        if( (src = cvLoadImage(path,-1)) == 0 )
        {
            printf("Cannot load file image %s\n", path);
            exit(0);
        }
        int w = src->width;
        int h = src->height;
        int sbin = 8;
        int interval = 10;
        double scale = pow(2., 1./interval);
        int m = (w<h)?w:h;
        int max_scale = 1+floor((double)log((double)m/(5.*sbin))/log(scale));
        image *ims = calloc(max_scale+interval, sizeof(image));

        for(i = 0; i < interval; ++i){
            double factor = 1./pow(scale, i);
            double ih =  round(h*factor);
            double iw =  round(w*factor);
            int ex_h = round(ih/4.) - 2;
            int ex_w = round(iw/4.) - 2;
            ims[i] = features_output_size(net, src, ex_h, ex_w);

            ih =  round(h*factor);
            iw =  round(w*factor);
            ex_h = round(ih/8.) - 2;
            ex_w = round(iw/8.) - 2;
            ims[i+interval] = features_output_size(net, src, ex_h, ex_w);
            for(j = i+interval; j < max_scale; j += interval){
                factor /= 2.;
                ih =  round(h*factor);
                iw =  round(w*factor);
                ex_h = round(ih/8.) - 2;
                ex_w = round(iw/8.) - 2;
                ims[j+interval] = features_output_size(net, src, ex_h, ex_w);
            }
        }
        for(i = 0; i < max_scale+interval; ++i){
            image out = ims[i];
            //printf("%d, %d\n", out.h, out.w);
            fprintf(fp, "%d, %d, %d\n",out.c, out.h, out.w);
            for(j = 0; j < out.c*out.h*out.w; ++j){
                if(j != 0)fprintf(fp, ",");
                fprintf(fp, "%g", out.data[j]);
            }
            fprintf(fp, "\n");
            free_image(out);
        }
        free(ims);
        fclose(fp);
        cvReleaseImage(&src);
        n = n->next;
    }
}

void features_VOC_image(char *image_file, char *image_dir, char *out_dir)
{
    int i,j;
    network net = parse_network_cfg("cfg/voc_features.cfg");
    char image_path[1024];
    sprintf(image_path, "%s%s",image_dir, image_file);
    char out_path[1024];
    sprintf(out_path, "%s%s.txt",out_dir, image_file);
    printf("%s\n", image_file);
    FILE *fp = fopen(out_path, "w");
    if(fp == 0) file_error(out_path);

    IplImage* src = 0;
    if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path);
    int w = src->width;
    int h = src->height;
    int sbin = 8;
    int interval = 10;
    double scale = pow(2., 1./interval);
    int m = (w<h)?w:h;
    int max_scale = 1+floor((double)log((double)m/(5.*sbin))/log(scale));
    image *ims = calloc(max_scale+interval, sizeof(image));

    for(i = 0; i < interval; ++i){
        double factor = 1./pow(scale, i);
        double ih =  round(h*factor);
        double iw =  round(w*factor);
        int ex_h = round(ih/4.) - 2;
        int ex_w = round(iw/4.) - 2;
        ims[i] = features_output_size(net, src, ex_h, ex_w);

        ih =  round(h*factor);
        iw =  round(w*factor);
        ex_h = round(ih/8.) - 2;
        ex_w = round(iw/8.) - 2;
        ims[i+interval] = features_output_size(net, src, ex_h, ex_w);
        for(j = i+interval; j < max_scale; j += interval){
            factor /= 2.;
            ih =  round(h*factor);
            iw =  round(w*factor);
            ex_h = round(ih/8.) - 2;
            ex_w = round(iw/8.) - 2;
            ims[j+interval] = features_output_size(net, src, ex_h, ex_w);
        }
    }
    for(i = 0; i < max_scale+interval; ++i){
        image out = ims[i];
        fprintf(fp, "%d, %d, %d\n",out.c, out.h, out.w);
        for(j = 0; j < out.c*out.h*out.w; ++j){
            if(j != 0)fprintf(fp, ",");
            fprintf(fp, "%g", out.data[j]);
        }
        fprintf(fp, "\n");
        free_image(out);
    }
    free(ims);
    fclose(fp);
    cvReleaseImage(&src);
}

int main(int argc, char *argv[])
{
    //feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);

@@ -406,7 +568,9 @@
    //test_ensemble();
    //test_nist();
    //test_full();
    train_VOC();
    //train_VOC();
    features_VOC_image(argv[1], argv[2], argv[3]);
    printf("Success!\n");
    //test_random_preprocess();
    //test_random_classify();
    //test_parser();