[CVPR2022 Oral] The official code for "TransRank: Self-supervised Video Representation Learning via Ranking-based Transformation Recognition"
Since I'm busy with some new projects now, I may not be able to release a clean implementation with sufficient guidance right now. But I promise to do it in late 2022.
Besides, if you are interested to check our implementation now, you can refer to the selfsup branch of my mmaction2 fork. It includes all the implementation of this paper, but may be a little bit messy. I also provide an example config: transrank_ucf_r3d_rotate_mlp_200e.py, which trains TransRank with 3 temporal transformations and 1 spatial transformation. Once you have cloned the repo and switched to the selfsup branch, you can run training with following command (8 GPUs):
bash tools/dist_train.sh configs/transrank_ucf_r3d_rotate_mlp_200e.py 8Also, please send me an email to me or open a new issue under this repo if you have any questions regarding to the implementation. My email is dhd.efz@gmail.com.