HYPJUDY / Decouple-SSAD

Decoupling Localization and Classification in Single Shot Temporal Action Detection
https://arxiv.org/abs/1904.07442
MIT License
96 stars 19 forks source link

window size of activitynet #3

Closed 666zcli closed 5 years ago

666zcli commented 5 years ago

hi,thanks for you excellent work. I found that an issue mentioned that you had some experiments on activitynet. Due to the differences between duration of videos in activitynet. So could you tell me what window size you use in your experiment on activitynet? Thanks

HYPJUDY commented 5 years ago

Hi, As mentioned in https://github.com/HYPJUDY/Decouple-SSAD/issues/2 "ActivityNet has larger scale in video length and action instance length (can be very short and very long).“ It's hard to set a proper window size which can cover most of the action instances for Anet, so I resize all videos to the same length (256 or 512) and use more (7 or 9) anchor layers. Maybe this manner of resizing videos will loss precision but BSN gets good performance with resized videos.

If you don't want to resize videos, then maybe you can try large window size like 1024?

666zcli commented 5 years ago

Thanks for your advice!