vid_demo.py after running, detected video as same as the original video.

[ZhangXiangYunfs]

Thanks for your jobs!
When I running the vid_demo.py, as your demo and example command ,the output video as same as the original video. I can't find where highlight the detected objects, and the created frames have no different with the video frames.
Hoping for your apply!

[YuHengsss]

Hi, could you please provide the concrete commands, the video used for inferencing to help us find the problem?

[ZhangXiangYunfs]

Hi!
My commands are as follows:
python tools/vid_demo.py --conf 0.25 --nms 0.5 --tsize 576 --save_result True
I modified the paths in the file, including the video path, weights path and the v++_SwinBaseX_decoupleReg.py's path under exps.
I choice the swinbase weights.
The original video is 30 seconds long and exceeds 10MB, so I sliced ​​a portion of the video.

video_7244810666815524151_10_76700_79670.mp4

Thanks for your apply!!

[YuHengsss]

I think the problem here is that the provided weights here is trained on ImageNet VID dataset, and the it's category does not contain "human". You can find the supported classes here "

YOLOV/yolox/data/datasets/vid_classes.py

Line 1 in 2ea4eb9

VID_classes = (

"

[ZhangXiangYunfs]

Thank you very much for your reply！
I hope I can detect it normally after training with my datasets.
In addition, could you provide guidance and instructions on the training preparation and training process in the Readme? This would be very useful to me.
I wrote "apply" instead of "reply" earlier. Sorry about that... haha...

[YuHengsss]

Thank you for your valuable suggestion. The problem of "Training on customed dataset" is a frequently asked question and we will add this part in the near feature. For now, you may find some answers in previous issues.

[ZhangXiangYunfs]

Get it，thanks again！