Self-Supervised Learning of Structured Dynamics from Videos
2026-07-23 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors study how to separate camera movement from object movement when watching videos, which is hard because these motions often mix together. They introduce a method called the Structured Dynamics Model (SDM) that uses features from a pretrained image model to predict future video frames while keeping these motion types separate. Their approach combines learning from real and simulated videos and performs better than other models on tests involving different kinds of motion. This work shows that existing image models can be adapted to better understand how things move in videos without needing heavy supervision.
motion representationcamera motionobject motionvision transformerfuture feature predictionself-supervised learningsynthetic datavideo dynamicslatent representationinductive bias
Authors
Lukas Knobel, Andrew Zisserman, Yuki M. Asano
Abstract
Understanding motion in video is a fundamental challenge for visual learning, as frame-to-frame change entangles two sources of dynamics: camera motion and object motion. This decomposition has remained underexplored in representation learning, partly because these factors are tightly coupled in natural videos and difficult to supervise separately. Yet recovering it is important for learning robust motion representations that separate meaningful object dynamics from camera-induced variation. We study whether such structured motion representations can be recovered from frozen features of a pretrained image vision transformer. We propose the Structured Dynamics Model (SDM), which explicitly separates the dominant source of temporal change from residual dynamics through future-feature prediction, rather than representing video change with a single entangled latent or with unstructured, spatially dense transition tokens. Training combines self-supervised learning on real video with weak supervision of scene dynamics on synthetic Kubric data. We evaluate SDM on ProbeMotion, a new evaluation suite spanning synthetic and real videos with camera motion, object motion, and combined dynamics. SDM outperforms backbone baselines using global CLS or average-pooled features, and compares favorably to strongly supervised representations such as VGGT on several probes, despite using substantially weaker supervision. These results suggest that pretrained image models can be readily repurposed into structured video-dynamics representations, providing a useful inductive bias for learning and analyzing latent video dynamics.