An Improved Inter-Intra Contrastive Learning Framework on Self-Supervised Video Representation

In this paper, we propose a self-supervised contrastive learning method to learn video feature representations. traditional methods, constraints from anchor, positive, and negative data pairs are used train the model. such case, different samplings of same treated as positives, clips videos negatives. Because spatio-temporal information is important for representation, set temporal more strictly by introducing intra-negative samples. addition samples videos, extended breaking relations in anchor video. With proposed Inter-Intra Contrastive (IIC) framework, can convolutional networks representations videos. Strong augmentations, residual clips, well head projector utilized construct an improved version. Three kinds generation functions extensive experiments using network backbones conducted on benchmark datasets. Without pre-computed optical flow data, our version outperform previous IIC large margin, 19.4% (from 36.8% 56.2%) 5.2% 15.5% 20.7%) points improvements top-1 accuracy UCF101 HMDB51 datasets retrieval, respectively. For recognition, over 3% also be obtained these two Discussions visualizations validate that IICv2 capture better clues indicate potential mechanism.