Motion-Centric Self-Supervised Video Representation Learning

A new arXiv paper (2605.23045) proposes a self-supervised video representation learning method that uses motion as the central modality. The approach employs point-tracks to capture motion and a masked-autoencoder to reconstruct missing tracks. This method avoids the high costs of scaling video models and the limitations of language-supervised learning, which restricts concepts to those in captions. The authors argue that current video models still struggle with temporal understanding, and their motion-focused technique aims to address this gap. The paper demonstrates that learning from motion alone can produce effective video representations without relying on language or large-scale datasets.