Reviving Unsupervised Optical Flow: Concept Reevaluation, Multi-Scale Advances and Full Open-Source Release

Unsupervised optical flow approaches have become more popular in the last decade, enabling the training of models across domains without ground truth data.Although RAFT and its successors have achieved significant success in supervised optical flow, many unsupervised approaches continue to use older backbones such as PWC-Net. One reason for this architectural stagnation is that SMURF, the current unsupervised RAFT-based SOTA, is not fully accessible and has proven challenging for the community to reproduce.In this paper, we revive and advance unsupervised optical flow: First, we introduce Un-RAFT, a simple yet effective unsupervised method with RAFT as backbone. Second, building on Un-RAFT, we present Un-CEMA: a novel unsupervised context-enhanced recurrent multi-scale approach, where context features are utilized for refining the flow, further improving both accuracy and preservation of details. Third, we reexamine previously advised unsupervised strategies to identify effective training settings. Finally, we fully open-source our Pytorch code publicly, enabling further advancements in the field. In terms of results, both our methods demonstrate strong generalization capabilities and set a new state of the art for unsupervised two-frame approaches on MPI-Sintel.Specifically, Un-CEMA surpasses SMURF with improvements up to 28%.