Self-Supervised Pre-Training for Optical Flow Estimation via Contrastive Learning

To address the issues of dataset dependency and correlation volume redundancy in optical flow estimation, contrastive learning is introduced to build a self-supervised optical flow framework. Firstly, adapting contrastive learning into optical flow to enable pre-training on open-source data; Secondly, using correlation volume as input instead of images, which producing better 2-D motion vectors; Finally, a Pretext Task based on motion aggregation is designed to generate both positive samples and query samples, and the expired positive embedding is stored into the negative sample queue. The model is first pretrained on YouTube-VOS, and then fine-tuned on optical flow datasets. EPE reaches 0.59 on the Sintel Clean sequence, with inference time of 76 milliseconds, exceeding both accuracy and efficiency of common optical flow estimation networks.