search

stevenlsw / hoi-forecast

[CVPR 2022] Joint hand motion and interaction hotspots prediction from egocentric videos
MIT License
54 stars 9 forks source link

Struggling to run rulstm feature extraction #17

Open Gregz9 opened 1 month ago

Gregz9 commented 1 month ago

Hi,

As part of my master thesis, I was hoping to run your model on some custom data. However, I am not able to get the environment for rulstm FasterRCNN setup correctly, and the model fails with CUDNN_STATUS_NOT_INITIALIZED. Do you possibly have any docker image available in which the feature extraction stage could be run?

stevenlsw commented 1 month ago

Hi,

Thanks for your interest. For feature extraction, basically you could follow Feature Extraction in RULSTM, for hand bbox detection you could use hand_object_detector to get hand and object bboxes, then you could use ROI to crop the hand and object features accordingly. For CUDNN_STATUS_NOT_INITIALIZED, i guess it is because you have not setup the cuda successfully. Can you check your current pytorch could use cuda like tensor.cuda()

Gregz9 commented 1 month ago

I managed to fix the CUDNN issue! Thank you for pointing me to hand_object_detector!