ATGS: Anchored Temporal Gaussian Splatting for Long Volumetric Video
janusch_patas · x · 2026-09-01
Presents ATGS, a Gaussian Splatting based framework for volumetric video reconstruction.
- Core Innovation: Instead of tracking individual Gaussian primitives for long-term complex motion, ATGS organizes Gaussians around time-conditioned anchors. This localizes their spatial and temporal support via anchors and a temporal windowing strategy, reducing motion complexity.
- Stability: Introduces a compact set of multi-level anchor features encoding global, local spatial, and local temporal features to jointly constrain Gaussian generation, ensuring spatial and temporal stability.
- Performance: Extensive experiments show ATGS consistently outperforms prior methods on long sequences with complex motions in terms of visual quality and temporal coherence.
More from Multimodal
- Merzmensch interview: Exploring human-machine memory via local AI — Merzmensch · 2026-09-01
- Single-word prompt series: generating Midjourney art from the Middle English word 'Selcouth' — tisch_eins · 2026-09-01
- ECCV 2026 Tutorial: Post-Training Alignment and Enhancement for Diffusion Models — RisingSayak · 2026-09-01
- Clip Bot: Extract Captioned Highlights from YouTube Videos Automatically — CodeByPoonam · 2026-09-01
- Worldfold adds volumetric fog with light scattering, runs on M1 Mac — jamestagg · 2026-09-01
- Runway Omni 1.1 Flash hands-on: Enhanced control with first/last frames — jnack · 2026-09-01