Closed GotG closed 4 years ago
Hi, @GotG Darknet:https://github.com/pjreddie/darknet/issues/427 Tensorflow(my project is based on mystic123 Tensorflow YOLOV3):https://github.com/mystic123/tensorflow-yolo-v3/issues/2 OpenVINOYOLOV3:https://docs.openvinotoolkit.org/latest/openvino_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html `python3 mo_tf.py --input_model /path/to/yolo_v3.pb --tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3.json --batch Since YOLOV3 ,OpenVINO didn't use --reverse_input_channels. I checked object_detection_demo_yolov3_async.py and i didn't found this demo converts BGR(OpenCV) to RGB(darknet and tf) So i did some tests: YOLOV4 original: YOLOV4 channels reversed: YOLOV4 tiny original: YOLOV4 tiny channels reversed:
I also do some video tests,the results are the same. Using --reverse_input_channels will be better So i agree with you. My conclusion is that this may be a mistake made by OpenVINO since YOLOv3
Thanks a lot!!
Thanks for checking. My results are also very close when tested on yolov3. I did mention this at some point on their forums.
Great job!! Thanks a lot!
Well done with your repo! Very useful! Do you happen to know for tiny yolo v4 which function prevents it from working in OpenVINO 20.01 or .02? Thanks
@GotG yolov4 tiny need this :
[route] layers=-1 groups=2 group_id=1
In Tensorflow,we could use https://github.com/TNTWEN/OpenVINO-YOLOV4/blob/534cca38d1e3bf95dce98daa216a57f62a2f83ce/yolo_v4_tiny.py#L21 or https://github.com/TNTWEN/OpenVINO-YOLOV4/blob/534cca38d1e3bf95dce98daa216a57f62a2f83ce/yolo_v4_tiny.py#L22 But they are not supported by OpenVINO before 2020R4(In fact, I used the 2019 version before ,so the exact version which began to support v4 tiny may not be accurate) Now OpenVINO 2020R4 supports the second operation but it still doesn't support the first operation.But anyway, that's enough for us to implement OpenVINO v4tiny
Hello
When converting to xml wondering whether reverse input channel is needed?
In my experience, tf models do require this.
Thanks