Peking University's DynStream does streaming 4D Gaussian reconstruction without camera poses
jiqizhixin · x · 2026-10-28
Peking University, VGI Labs and UC Merced introduce DynStream, a streaming 4D Gaussian reconstruction framework for videos where both camera and objects move, requiring no supplied camera poses.
- Bottleneck: local reconstructions drift in position and scale; moving objects leave ghosts if treated as permanent scene parts
- Mechanism: processes video in temporal chunks, predicts relative poses and local scale, stitches chunks via overlapping keyframes; static Gaussians accumulate in a shared scene while dynamic Gaussians stay timestamped, so new frames extend the reconstruction without re-optimizing from scratch
- Results: leading rendering quality across five datasets, including KITTI's 11 odometry sequences
More from Research
- Can AI search automatically discover cross-field math connections like Fermat's proof? — IgorCarron · 2026-10-28
- New bioRxiv paper reframes enzyme function prediction as conditional generation — KevinKaichuang · 2026-10-28
- New generative model predicts enzyme reaction mechanisms straight from sequence — KevinKaichuang · 2026-10-28
- GenAI is creeping into bestselling novels, study featured in Nature — TuhinChakr · 2026-10-28
- Nathan Lambert: open models' demise exaggerated, but AA index flattered them — natolambert · 2026-10-28
- Local NCM benchmarks: 4x faster for one request, 4-8x slower at 64 concurrent — Prompt Engineering · 2026-10-28