When webcam_demo.py and long_video_demo.py perform motion recognition on the same video, their results are much different.

[Air847735]

I fine-tuned tsn using my own dataset during experiments. The results I got are working fine in webcam_demo. But I wanted to be able to identify each picture of the video, so I tried long_video_demo.py, but I found that the same video was shot on the screen through the webcam and the video was directly identified, the results were very different. , but the results of Webcam are closer to correct, while long_video_demo cannot get good results, and even recognizes errors. My config and label_map both use the same one. Does anyone know what could cause this to happen?