Self-Supervised Learning of Structured Dynamics from Videos

2026-07-23

AI Summary

This work addresses the challenge of disentangling camera motion from object motion in video frames—a key obstacle to learning robust dynamic representations. The authors propose Structured Dynamic Modeling (SDM), a paradigm that leverages frozen pretrained Vision Transformer (ViT) features to explicitly separate dominant dynamics from residual dynamics through future feature prediction. SDM is trained using a combination of self-supervised signals from real videos and weakly supervised signals from synthetic Kubric data, enabling it to extract disentangled dynamic representations from static image models without strong supervision. This approach overcomes limitations of conventional methods that rely on single latent variables or unstructured transition modeling. Evaluated on the newly introduced ProbeMotion benchmark suite, SDM substantially outperforms CLS token or average pooling baselines and matches the performance of the strongly supervised VGGT model across multiple probing tasks, demonstrating its effectiveness.

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.

Problem

• structured dynamics
• self-supervised learning
• video motion decomposition
• camera motion
• object motion

Innovation

• Self-Supervised Learning
• Structured Dynamics
• Video Representation
• Future-Feature Prediction
• Vision Transformer