discussion and implementation details. 23 [ All Reviewers ] Related work. We agree with the reviewers that a more extended discussion is required for related

This model is trained on UCF101 with the same schedule as CoCLR for a fair comparison. We note a recent arXiv paper (MemDPC, to appear in ECCV2020 by Han et al.) has also used both RGB and optical We will add these discussions. We actually used the same augmentation as DPC in their released codebase. These are the core contributions of our paper. CoCLR-RGB model gets 70.2% by linear probing.