A recent arXiv paper investigates how frozen pretrained visual and depth representations can be reused for dense optical flow. The authors build on TTSFlow, a prior optical flow architecture, and study the Dense Prediction Transformer (DPT) feature space as a source of frozen representations.
The core idea is to avoid fine-tuning the pretrained backbone and instead attach a supervised matching network that maps these features to flow estimates. This approach tests whether general-purpose visual and depth features carry enough geometric and motion information for accurate dense correspondence.
The abstract is brief and does not yet report quantitative results, so the main contribution at this stage is the architectural direction. The paper appears to be a cross-listed announcement, meaning the full details may be available elsewhere, but the provided text only outlines the method and its lineage from TTSFlow.