In this part, we give implementation of 3D-Resnet with Full-Conv. 3D-Resnet implementation is built from Kensho Hara's repository.
Please follow the steps in that repository for requirements and arranging the dataset.
We have 2 different implementations of Full-Conv:
1. Full-Conv on spatial dimension.
python main.py --root_path <your path> --video_path <your video path> --annotation_path <your annotation path/ucf101_01.json> --result_path <your path for results> --dataset ucf101 --model resnet --model_depth 18 --n_classes 101 --batch_size 32 --n_threads 4 --checkpoint 5
2. Full-Conv on spatial and temporal dimension. After the submission, we extended the Full-Conv on temporal dimension and the new model obtained an extra 1.6% increase with 3D-Resnet18 arhitecture on UCF101 dataset.
python main.py --root_path <your path> --video_path <your video path> --annotation_path <your annotation path/ucf101_01.json> --result_path <your path for results> --dataset ucf101 --model resnet --model_depth 18 --n_classes 101 --batch_size 32 --n_threads 4 --checkpoint 5 --temporal True
Note: If you want to use a network with Same-Conv, then use --f_conv False argument.