Peking University's DynStream does streaming feed-forward 4D Gaussian reconstruction
机器之心 · wechat · 2026-10-19
A team from Peking University's Wangxuan Institute (Prof. Yongtao Wang) with VGILabs and UC Merced presents DynStream, an online streaming 4D Gaussian feed-forward reconstruction framework: given a pose-free RGB video stream, it outputs camera trajectories and continuously updatable dynamic scene representations without per-scene optimization. Paper: arxiv.org/pdf/2610.09720
Challenges addressed: incremental processing, global consistency over time, high-fidelity dynamic scene representation.
Three core modules
- Local online reconstruction: video is chunked temporally; a causal-attention backbone feed-forwardly reconstructs local scenes, avoiding attention costs that explode with video length.
- Computation-free global aggregation: pose prediction is reformulated as relative poses plus a predicted scale factor mapping to metric scale; overlapping keyframes align chunks to a global coordinate frame, eliminating costly SIM(3) optimization.
- Static-dynamic separation: each Gaussian gets a dynamic probability; static Gaussians join the global scene while dynamic ones render per timestep, suppressing ghosting.
Results
- Best on all rendering metrics across 5 datasets (Waymo, SpatialVID, etc.) vs AnySplat, MVSplat, NeoVerse; cleaner edges and fewer dynamic ghosts.
- On 11 KITTI sequences (23,201 frames, up to 5km), Sim(3) ATE drops from hundreds of meters to 40m.
Demos include a 2,206-meter driving scene. First author is PKU undergrad Xian Dingwei.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23