UniQuery4R encodes a clip once and unifies 4D scene reconstruction via query-conditioned decoding
zhenjun_zhao · x · 2026-08-19
New arXiv paper (2608.17283, Tiancheng Chen et al.). Dynamic 4D scene reconstruction requires jointly estimating correspondence, geometry, object motion, and camera motion; existing feed-forward methods predict dense task-specific maps or process source-target pairs independently, wasting computation on sparse queries and limiting feature reuse.
UniQuery4R is a query-conditioned framework: a multi-frame clip is encoded once, and source view, target view, and continuous source-image coordinates are selected only at decoding time via source-to-target cross-attention. Each query jointly predicts target correspondence, target-time 3D position, scene flow, and source depth, with camera parameters estimated per view. The encoded clip is reused across arbitrary source-target selections, supporting both sparse inference and dense reconstruction through batched queries without temporal embeddings tied to a fixed clip length. A direction-magnitude parameterization of scene flow with separate supervision for moving vs. static points is introduced. UniQuery4R achieves the best macro-average results on WorldTrack for both scene-flow estimation and dynamic-point reconstruction.
More from Multimodal
- From character creation to stories: AI influencers streamline anime workflows — aftahi_ai · 2026-08-20
- Grok demo: Accurate language accent and long 1080p video generation — elonmusk · 2026-08-20
- AI Tool GeoSpy Locates Photos from Pixels with Meter-Level Accuracy — saibharadwaj · 2026-08-20
- Runway Gen-2 Update: 1080p Support, 50 References, 30s Generation via API — tlakomy · 2026-08-20
- Digital Sculpting: Creating the Thesis Rock with Rendering Magic — every · 2026-08-20
- Swarms Builds Inference Engines for Media Generation without Frameworks — bingxu_ · 2026-08-20