Contrastive Transformation for Self-supervised Correspondence Learning

In this paper, we focus on the self-supervised learning of visual correspondence using unlabeled videos in wild. Our method simultaneously considers intra- and inter-video representation associations for reliable estimation. The intra-video transforms image contents across frames within a single video via frame pair-wise affinity. To obtain discriminative instance-level separation, go beyond analysis construct affinity to facilitate contrastive transformation different videos. By forcing consistency between levels, fine-grained are well preserved feature discrimination is effectively reinforced. simple framework outperforms recent methods range tasks including object tracking (VOT), segmentation (VOS), pose keypoint tracking, etc. It worth mentioning that our also surpasses fully-supervised (e.g., ResNet) performs competitively against algorithms designed specific VOT VOS).