Depth Anything V4: Dynamic 4D Scene Reconstruction via Riemannian Flow Matching on 4D Gaussian Splatting
- Published
- Source
- arXiv
- Paper number
- 951
- Field
- Computer Vision
- arXiv ID
- 2608.18388
Key points
- Gaussian-splatting parameters such as scale, rotation, and opacity lie on curved manifolds, so linear interpolation in Euclidean space can produce invalid intermediate values. Interpolation on the manifolds keeps every intermediate state valid.
- With the data, architecture, and test-time optimization held constant, the deterministic MLP achieved an F-score of 0.762, while Riemannian Flow Matching reached 0.806. The 0.044 gain isolates the contribution of flow matching.
- In the Euclidean-space variant, 12.3% of the Gaussians took invalid values, and the F-score fell to 0.684.
- Its uncertainty calibration was substantially better than MC Dropout, with NGLL of -1.58 versus -0.82 and ECE of 0.058 versus 0.18. Its confidence therefore tracked both certainty and uncertainty more reliably.
- It outperformed earlier Depth Anything models and per-scene optimized 4D-GS without using human-annotated depth labels. The 360 GPU-hour pretraining cost becomes lower than per-scene optimization when deployed across more than 10,000 scenes.
Paper links
External research summaries. These are not HDATF publications or measured product results.