search

joslefaure / HIT

Official Implementation of our WACV2023 paper: “Holistic Interaction Transformer Network for Action Detection”
https://arxiv.org/abs/2210.12686
55 stars 9 forks source link

Could MultiSports Dataset be tranformed to like ucf101-24 dataset or JHMDB dataset? #27

Open Eurususu opened 1 year ago

Eurususu commented 1 year ago

as showed in title

joslefaure commented 1 year ago

It could. To get started with Multisports, I would suggest the other repository HIT_ava. The code does not follow the methodology of the published paper but it's simpler and more accurate (at least for UCF and JHMDB), while requiring only person annotations.

Eurususu commented 1 year ago

Thanks for suggestion. I have tried convert Multispots to AVA format dataset followinghttps://github.com/Whiffe/Custom-ava-dataset_Custom-Spatio-Temporally-Action-Video-Dataset, but I extracted key frame every five frames. When I test the custom dataset on SlowFast, it was a shit! (and in Multisports some actions just few frames) So I wanna ask how do you make custom AVA dataset and how many frames should be skipped to extract a keyframe?

joslefaure commented 1 year ago

I did not skip any frame for AVA (ran the code on video data) For Multisports, I use all the frames provided by the authors of the dataset.

ddddqt commented 6 months ago

@Eurususu Hello, can you provide the code to convert multisports into ava data format and train in hit_ava?thank you very much