search

MCG-NJU / MultiSports

[ICCV 2021] MultiSports: A Multi-Person Video Dataset of Spatio-Temporally Localized Sports Actions
Other
103 stars 7 forks source link

Baseline and MOC-Detector #1

Open zhufeng888 opened 3 years ago

zhufeng888 commented 3 years ago

I have many questions as follows: ①when the baseline will be released and what model it will be based on? ②I'm trying to use MOC-Detector to reproduce the result you provide, but I don't know how do you generate optical flow images. The MOC-Detector uses puppet flow(it is hard for me to understand. Are puppet flow images generated by calibrated manually?). ③Finally, I put the rawframe generated by your generate_rgb.py to MOC-Detector ,but one epoch consums over 5 hours in 8 tesla p100 gpu(also one of the reason is that my cpu is too slow to load data).MultiSports Data is over 60G while is JHMDB below 10G? Is there any suggestion for the contestants to speed up the training. I hope to get your answer or advice. Than you and wish you a happy.

yixuanli98 commented 2 years ago

We do not use the flow for multisports.

zhufeng888 commented 2 years ago

We do not use the flow for multisports.

Thanks for your reply. I meet another issue on submission, and I have created a new issue. Could you help me？ https://github.com/MCG-NJU/MultiSports/issues/2

yixuanli98 commented 2 years ago

What's more, you can comment this line https://github.com/MCG-NJU/MOC-Detector/blob/45c3fd0c5e4ca455e7e739f3ad80b2c91813df35/src/datasets/sample/sampler.py#L62 to speed up, which is a time-consuming but not important data augmentation.

yixuanli98 commented 2 years ago

We will release both the MOC and slowfast codes and models.