Across 1 000 hours of video from 22 cities, FactorJEPA demonstrates that decomposing future states into layout, entity, and interaction channels slashes the amount of supervision required while still capturing the chaotic dynamics of Global‑South traffic [1].
Prior world‑model work on traffic has often used monolithic latent predictions built atop V‑JEPA backbones and typically relies on dense frame‑wise annotations to model heterogeneous agents and persistent occlusions.
FactorJEPA improves future‑latent accuracy, intervention‑sensitive prediction, and robustness to reduced visual evidence, achieving a Spearman ρ between 0.895 and 0.978 across the 1 B and 2 B backbones
To guarantee that gains are not artefacts of label leakage, the authors “strictly separate training‑target construction from evaluation” — yet the reported motion‑information trade‑off remains only characterized by a Motion cosine metric, leaving open how this balance scales with even larger scene graphs [1].
The released DENSEWORLD‑115k dataset of 115 k annotated scenes provides a benchmark for evaluating FactorJEPA, and its availability may reduce the need for extensive labeling.
Top comments (0)