StereoSplat: metric-scale novel view synthesis from stereo, AbsRel 0.03
zhenjun_zhao · x · 2026-10-29
StereoSplat is a feed-forward 3DGS architecture built for stereo video, achieving reliable metric-scale novel view synthesis. It extracts dense multi-scale features and disparity from a FoundationStereo backbone, decodes per-pixel Gaussians via cross-view fusion and gated depth refinement, and clusters primitives with learnable depth-adaptive aggregation.
- Trained purely on synthetic data (new XRStereo dataset), transfers zero-shot to real stereo video
- On Infinitigen: depth AbsRel 0.03, δ1 98.6, PSNR 29.4, far ahead of DepthSplat (0.21/63.0)
- Also leads on TartanV2 and KITTI without ground-truth poses
- Adaptive voxelization trades compression rate against quality
More from Research
- AlphaBet: agentic RL trains LLM forecasters, cutting Brier score to 0.626 — GeorgiaTech · 2026-10-29
- Information-theoretic scaling laws predict memorization and FID in diffusion models — LucaAmb · 2026-10-29
- ICLR reviewer hands out scores of 1,1,1,1 and 2 across five papers — chhaviyadav_ · 2026-10-29
- kalomaze calls out hnet baseline: effective depth unmatched, no wall clock reported — kalomaze · 2026-10-29
- AVERT Study Launched: Can Autonomous Vehicles Reduce Trauma? — zakkohane · 2026-10-29
- Bad binder design drove only ~3% of Phase 2/3 failures — the rest elude AlphaFold — clarejtbirch · 2026-10-29