Single-channel YOLOv4 Tiny inference with use OpenVINO

[susanin1970]

Hi!
I have trained on single-channel images YOLOv4 TIny .weights file

Then I got frozen_darknet_yolov4_model.pb file of this model with use follow command:

python .\convert_weights_pb.py --class_names "D:\Repositories\isocontainercodes-training\pipeline\yolo\names.list" --weights_file "D:\Repositories\isocontainercodes-training\pipeline\yolo\yolov4-tiny-mixed_y8-v3.weights" --data_format NHWC --tiny

Before that, I changed the line in convert_weights_pb.py so that there were not 3 channels, but 1:

    # placeholder for detector inputs
    inputs = tf.placeholder(tf.float32, [None, FLAGS.size, FLAGS.size, 1], "inputs")

Then I got .bin and .xml files of OpenVINO IR model with use following command:

"C:\Program Files (x86)\Intel\openvino_2021.4.582\deployment_tools\model_optimize\mo.py" --input_model .\frozen_darknet_yolov4_model.pb --transformations_config .\yolo_v4_tiny.json --batch 1 --reverse_input_channels

When I tried to run inference:

python .\pythondemo\2021.3\object_detection_demo_yolov3_async.py -i "C:\Users\Reutov\Desktop\test_nmtp_video_4.mp4" -m .\frozen_darknet_yolov4_model.xml -d CPU

I got this:

Before that I changed code of object_detection_demo_yolov3_async.py:

def resize(image, size, keep_aspect_ratio, interpolation=cv2.INTER_LINEAR):
    if not keep_aspect_ratio:
        return cv2.resize(cv2.cvtColor(image, cv2.COLOR_BGR2GRAY), size, interpolation=interpolation)

    iw, ih = image.shape[0:2][::-1]
    w, h = size
    scale = min(w/iw, h/ih)
    nw = int(iw*scale)
    nh = int(ih*scale)
    image = cv2.resize(cv2.cvtColor(image, cv2.COLOR_BGR2GRAY), (nw, nh), interpolation=interpolation)
    new_image = np.full((size[1], size[0], 3), 128, dtype=np.uint8)
    dx = (w-nw)//2
    dy = (h-nh)//2
    new_image[dy:dy+nh, dx:dx+nw, :] = image
    return new_image  

# Read and pre-process input images
    if net.input_info[input_blob].input_data.shape[1] == 3:
        input_height, input_width = net.input_info[input_blob].input_data.shape[2:]
        nchw_shape = True
    else:
        input_height, input_width = net.input_info[input_blob].input_data.shape[2:]
        nchw_shape = False  

.weights model YOLOv4 Tiny works in darknet normally:

What should I do to make the YOLOv4 Tiny single-channel model give correct detections using OpenVINO?
Thanks in advance