Question about test

[wjn922]

Hi,

I ran the code of test, and the results are shown as blow.

I wonder whether the repo supports evaluation of mAP of iou@[03:0.7], as reported in the paper.
Another question is why the inference needs so much time?
Thanks.

[C-H-Wong]

When evaluating different mAP, you only need to pass different iou_thr value to evaluate function here in tools/test.py here. We suggest to save the inference results first (set --out when testing), then run evaluation with different iou_thr, since inference is quite time-consuming.

While comparing to training process which only deal with a clip of 768-frame-length per video, testing would process whole video. Since the model can only process limited length of sequential data, videos are cropped by an overlap ratio of 25% when testing. Besides, processing video data is quite time-consuming already. That's what makes testing process look much more slower than training. After our optimization, the speed of testing process is already lifted a lot.

[wjn922]

Thanks for your clarification!

[wjn922]

Since the inference needs so much time, will you support the DDP testing in the future?