Darknet: different results libdarknet.so and ./darknet detector test

[kebundsc]

@AlexeyAB
Hello and many thanks for this repository.
I am trying to build my own *.c application using libdarknet.so and so far everything works fine. But if I compare the results generated my application and the results generated by using ./darknet detector test ... i receive slightly different (smaller) bounding boxes. I am using the same *.cfg and *.weights files. I also tried to use the same parameter settings as in the default detector.c -> test_detecor(). Can somebody explain me whats exactly causing these different results?

this is my application src.

int predict(char *inputs) {

    // -----------------------------------------------------------------------------------------------------------------
    // Define constants that were used when Darknet network was trained.
    // This is pretty much hardcoded code zone, just to give an idea what is needed.
    // -----------------------------------------------------------------------------------------------------------------
    char *datacfg = "/home/je/Dokumente/darknetAlexey/darknet/cfg/ast.data";

    metadata metas = get_metadata(datacfg);
    char **names = metas.names; 

    // Path to configuration file.
    char *cfgfile = "/home/je/Dokumente/darknetAlexey/darknet/cfg/yolov3-ast_test.cfg";

    // Path to weight file.
    char *weightfile = "/home/je/Dokumente/darknetAlexey/darknet/backup/yolov3-ast_train_4900.weights";

    // Define thresholds for predicted class.
    float thresh = .5;
    float hier_thresh = .5;

    // -----------------------------------------------------------------------------------------------------------------
    //
    // Do actual logic of classes prediction.
    // -----------------------------------------------------------------------------------------------------------------

    // Load Darknet network itself.
    //network *net = load_network_custom(cfgfile,weightfile, 1, 1); // set batch=1
    network *net = load_network(cfgfile, weightfile, 0);
    //network net = *net_ptr

    //set_batch_network(net, 1);



    srand(2222222);
    char buff[256];
    char *input = buff;
    float nms = .45; // 0.4F
    int j;

    input = "/home/je/Dokumente/darknetAlexey_backup/darknet/data/BBox-Label-Tool-master/JEdevkit/2019/images/1554208528848.jpg";



    image im = load_image_color(input, 0, 0);
    //image sized = load_image_resize(input, net->w, net->h, net->c, &im);

    image sized = resize_image(im, net->w, net->h);


    //image sized = letterbox_image(im, net->w, net->h); //letterbox = 1;


    float *X = sized.data;

    double time = get_time_point();
    network_predict_ptr(net, X);
    //network_predict_image(&net, im); //letterbox = 1;

    printf("%s: Predicted in %lf milli-seconds.\n", input, ((double)get_time_point() - time) / 1000);
        //printf("%s: Predicted in %f seconds.\n", input, (what_time_is_it_now()-time));

    int nboxes = 0;
    detection *dets = get_network_boxes(net, im.w, im.h, thresh, hier_thresh, 0, 1, &nboxes,0);

    if (nms) do_nms_sort(dets, nboxes, 1, nms);

    printf("Darknet application\n im w: %i \n im h: %i \n",im.w,im.h);

    draw_detections_v3(im, dets, nboxes, thresh, names, 0,1,1);

    save_image(im, "predictions");
    show_image(im, "predictions");


    free_detections(dets, nboxes);
    free_image(im);
    free_image(sized);


    wait_until_press_key_cv();
    destroy_all_windows_cv();


    // free memory
    free_ptrs((void**)names, net->layers[net->n - 1].classes);

    free_network(net);

    return 0;
}

The following picture shows the extOuptut for both cases.

Thank You!
Kevin

[AlexeyAB]

@kebundsc Hi,

Try to use float thresh = .25; insted of float thresh = .5;

Use network *net = load_network_custom(cfgfile,weightfile, 1, 1); instead of network *net = load_network(cfgfile, weightfile, 0);

[kebundsc]

Thank You @AlexeyAB
I accidentally used two different .cfg files. So everything works perfect.
Whats the difference between the two mentioned ways of loading the network?

[AlexeyAB]

network *net = load_network(cfgfile, weightfile, 0); loads params batches= and timesteps= from cfg-file - is required for Training

network *net = load_network(cfgfile, weightfile, 1, 1); force params batches=1 and timesteps=1 - is required for Testing