fix to run on TF2.4 about tensorflow-2.x-yolov3 HOT 6 OPEN

@JuliusJacobitz With TF 2.3 everything works without any problems.

from tensorflow-2.x-yolov3.

@ZhengzhiLiu are you able to reproduce the same results in terms of speed on your machine? If yes, can you tell me how fast the detect_video is running and what GPU you are using? I am unable to reproduce the quoted results in terms of fps using pretrained yolov4/yolov3 model files

from tensorflow-2.x-yolov3.

@yugkhanna I was able to reproduce the reported about 20 FPS in the code README.md on both TF2.3 and TF2.4. In order to run the code on TF2.4, you may need to make the changes as I mentioned above. On TF2.4 detect_video runs at 20 FPS on my RTX3080 (only support TF2.4 and above for GPU version)

.

YOLO tiny is able to run at 55 FPS, but the result is bad, bbox is too small and does not cover the entire object, and sometimes a single object has multiple bboxes, it seems nms failed.

from tensorflow-2.x-yolov3.

@ZhengzhiLiu I also noticed like @yugkhanna that with this fix at detect_video the fps are lower than before.
I have tested with detection_demo and the test.mp4 file.

In comparison (Python 3.7, TF 2.3, GPU: 980ti):
yolo.predict(image_data) i get ~16fps
yolo(image_data) i get ~8fps

Do you have any idea what could be the reason for this?

I also noticed at TF >= 2.4 that when printing pred_bbox in detect_video with yolo.predict(image_data), the bounding boxes contain correct values in the first frame and then consist only of nan.

from tensorflow-2.x-yolov3.

I also noticed at TF >= 2.4 that when printing pred_bbox in detect_video with yolo.predict(image_data), the bounding boxes contain correct values in the first frame and then consist only of nan.

@Crytap I'm currently struggling with exactly this problem. Do you have any advice which tf version to use?

from tensorflow-2.x-yolov3.

@ZhengzhiLiu Hello. What other changes have you made to increase the frame rate to 21?

from tensorflow-2.x-yolov3.