Tiny-Yolov2 Detection Output

[Chakraborty__Subhasi]

I have tried to modify the python samples to run tiny-yolov2 through openvino. I have converted the tiny-yolov2 model and .weight file through Darkflow to .pb format and then used the documentation of Openvino R3 to successfully convert the .pb files to IR representation. But after modifying the source code the code runs successfully but there are no detections. I am attaching the sample code of post-processing, please throw some light over the issue

def sigmoid(p):
    return 1.0 / (1 + math.exp(-p * 1.0))


def overlap(x1, w1, x2, w2):  # x1 ,x2 are two box center x
    left = max(x1 - w1 / 2.0, x2 - w2 / 2.0)
    right = min(x1 + w1 / 2.0, x2 + w2 / 2.0)
    return right - left


def cal_iou(box, truth):
    w = overlap(box[0], box[2], truth[0], truth[2])
    h = overlap(box[1], box[3], truth[1], truth[3])
    if w < 0 or h < 0:
        return 0
    inter_area = w * h
    union_area = box[2] * box[3] + truth[2] * truth[3] - inter_area
    return inter_area * 1.0 / union_area


def apply_nms(boxes, thres):
    sorted_boxes = sorted(boxes, key=lambda d: d[7])[::-1]
    p = dict()
    for i in range(len(sorted_boxes)):
        if i in p:
            continue

        truth = sorted_boxes
        for j in range(i + 1, len(sorted_boxes)):
            if j in p:
                continue
            box = sorted_boxes
            iou = cal_iou(box, truth)
            if iou >= thres:
                p = 1

    res = list()
    for i in range(len(sorted_boxes)):
        if i not in p:
            res.append(sorted_boxes)
    return res


def post_processing(output):

    res = output.astype(np.float32)
    res=np.reshape(res,(13,13,125))

    swap = np.zeros((13 * 13, 5, 25))

    index = 0
    for h in range(13):
        for w in range(13):
            for c in range(125):
                i=h*13 + w
                j = int(c/25)
                k = c%25
                swap=res

    biases = [1.08, 1.19, 3.42, 4.41, 6.63, 11.38, 9.42, 5.11, 16.62, 10.52]

    boxes = list()
    for h in range(13):
        for w in range(13):
            for n in range(5):
                box = list();
                cls = list();
                s = 0;
                x = (w + sigmoid(swap[h * 13 + w][0])) / 13.0;
                y = (h + sigmoid(swap[h * 13 + w][1])) / 13.0;
                ww = (math.exp(swap[h * 13 + w][2]) * biases[2 * n]) / 13.0;
                hh = (math.exp(swap[h * 13 + w][3]) * biases[2 * n + 1]) / 13.0;
                obj_score = sigmoid(swap[h * 13 + w][4]);
                for p in range(20):
                    cls.append(swap[h * 13 + w][5 + p]);

                large = max(cls);
                for i in range(len(cls)):
                    cls = math.exp(cls - large);

                s = sum(cls);
                for i in range(len(cls)):
                    cls = cls * 1.0 / s;

                box.append(x);
                box.append(y);
                box.append(ww);
                box.append(hh);
                box.append(cls.index(max(cls)) + 1)
                box.append(obj_score);
                box.append(max(cls));
                box.append(obj_score * max(cls))
                # print("these are the values of box 5 and 6", box[5], box[6])
                # if score
                if box[5] * box[6] > 0.1:
                    boxes.append(box);
    res = apply_nms(boxes, 0.35)
    label_name = {0: "bg", 1: "aeroplane", 2: "bicycle", 3: "bird", 4: "boat", 5: "bottle", 6: "bus", 7: "car",
                  8: "cat", 9: "chair", 10: "cow", 11: "diningtable", 12: "dog", 13: "horse", 14: "motorbike",
                  15: "person", 16: "pottedplant", 17: "sheep", 18: "sofa", 19: "train", 20: "tvmonitor"}
    w = img_cv.shape[1]
    h = img_cv.shape[0]

    for box in res:

        xmin = (box[0] - box[2] / 2.0) * w;
        xmax = (box[0] + box[2] / 2.0) * w;
        ymin = (box[1] - box[3] / 2.0) * h;
        ymax = (box[1] + box[3] / 2.0) * h;
        if xmin < 0:
            xmin = 0
        if xmax > w:
            xmax = w
        if ymin < 0:
            ymin = 0
        if ymax > h:
            ymax = h

        cv2.rectangle(img_cv,(int(xmin),int(ymin)),(int(xmax),int(ymax)),(0,255,0),2)
        print (label_name[box[4]],xmin, ymin, xmax, ymax)
        
        label_text = label_name[box[4]] + " " + str("{0:.2f}".format(box[5]*box[6]))
        label_background_color = (70, 120, 70) # greyish green background for text
        label_text_color = (255, 255, 255)   # white text
        
        label_size = cv2.getTextSize(label_text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)[0]
        label_left = int(xmin)
        label_top = int(ymin) - label_size[1]
        label_right = label_left + label_size[0]
        label_bottom = label_top + label_size[1]


        cv2.rectangle(img_cv, (label_left-1, label_top-5),(label_right+1, label_bottom+1), label_background_color, -1)
        cv2.putText(img_cv, label_text, (int(xmin), int(ymin-5)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255,255,255), 1)
    cv2.imshow('YOLO detection',img_cv)
    cv2.waitKey(10000)

[Yuanyuan_L_Intel]

Please pay attention to the layout of the RegionYolo layer.  In your code, the assumed layout is 13x13x125 and converted to 13x13x5x25. But, in the openvino, the shape of output is <1, 21125> whose's layout is acutally 5x25x13x13.