RGB-Only Baseline for Temporal Action Detection

[AphroDatalyst]

I'd like to explore the feasibility of using the repository using only RGB images, without the dependency on flow kinetics data.
I'm open to hearing any feedback or suggestions you may have.

[Anirudh257]

You can modify the dataloaders and comment the motion/flow part and retrain the model again.

[zhaoyue-zephyrus]

Hi @ZohrehBayramalizadeh

Changing MODALITY: 'twostream' to MODALITY: 'visual' in the config will hopefully satisfy your request. Please see https://github.com/zhaoyue-zephyrus/TeSTra/blob/main/src/rekognition_online_action_detection/models/feature_head.py#L33 for all possible values. Feel free to re-open the comment if you still have any issue.

[Anirudh257]

Hi @zhaoyue-zephyrus there are some more changes needed in addition to the one suggested by you in the config and in the dataloaders part.

[Chenhongchang]

Hi @zhaoyue-zephyrus there are some more changes needed in addition to the one suggested by you in the config and in the dataloaders part.

HI! WHAT THE MORE CHANGES ARE? COULD YOU SHARE? @Anirudh257

[Anirudh257]

Hi @Chenhongchang change the dimensions.