This folder contains the training, finetuning code for the following modalities.
Pretrained models can be downloaded via Google Drive. Those pretrained models can be used to reproduce our final results using the testing code.
See requirements.txt
The RGB frames modality can be trained, finetuned and tested using the following commands in Conv3D/ folder.
python Sign_Isolated_Conv3D_clip.py
python Sign_Isolated_Conv3D_clip_finetune.py
python Sign_Isolated_Conv3D_clip_test.py
The Flow Color modality can be trained, finetuned and tested using the following commands
python Sign_Isolated_Conv3D_flow_clip.py
python Sign_Isolated_Conv3D_flow_clip_funtine.py
python Sign_Isolated_Conv3D_flow_clip_test.py
The HHA modality can be trained, finetuned and tested using the following commands
python Sign_Isolated_Conv3D_hha_clip_mask.py
python Sign_Isolated_Conv3D_hha_clip_mask_finetune.py
python Sign_Isolated_Conv3D_hha_clip_mask_test.py
The Flow Depth modality can be trained, finetuned and tested using the following commands
python Sign_Isolated_Conv3D_depth_flow_clip.py
python Sign_Isolated_Conv3D_depth_flow_clip_finetune.py
python Sign_Isolated_Conv3D_depth_flow_clip_test.py
The results .pkl files of the above modalities will be saved in results/ folder. Please rename and copy them to ../ensemble/ folder for model ensemble.