Peking University's ATGS Reconstructs 1,400-Frame Athletic Scenes in One Shot
A new Gaussian splatting framework reconstructs volumetric videos with thousands of frames of complex motion, extending prior methods by roughly 70x in temporal coverage.
PRO- ATGS reconstructs volumetric videos with thousands of frames of complex motion in a single training run.
- Time-conditioned anchors replace per-Gaussian trajectory tracking, avoiding the instability that breaks prior 4D methods.
- A temporal windowing strategy activates only anchors near the queried time to reduce jitter.
- Three feature levels (anchor, static spatial grid, local temporal hash grid) jointly constrain Gaussian generation.
- Beats LocalDyGS, 4DGaussian, and SpaceTimeGS on N3DV, MeetRoom, and VRU basketball benchmarks.
- Code available at github.com/WuJH2001/ATGS; offline pipeline requiring COLMAP initialization.
Volumetric video has a stubborn scaling problem. Most Gaussian splatting methods handle a couple of seconds of fast motion or a few hundred frames of gentle head-and-hand movement, but they fall apart on minute-long sequences with athletes running, jumping, and colliding. A team from Peking University and Pengcheng Laboratory has released ATGS (Anchored Temporal Gaussian Splatting), a framework that pushes single-shot reconstruction into the thousand-frame regime while outperforming prior methods on image quality.
Why long sequences break Gaussian splatting
Dynamic 3D Gaussian splatting works by either warping a canonical set of Gaussians over time (deformation-based methods) or placing 4D primitives directly in space-time. Frame-wise training handles long video but requires heavy storage and accumulates inter-frame errors that cause instability or model collapse in later stages. Deformation-based approaches struggle to track long-term dynamics, and recent 4D-Gaussian methods like FreeTimeGS and LocalDyGS handle complex motion only for one-to-two second clips.
The root cause is that explicitly tracking complex long-term motion with individual Gaussian primitives is inherently unstable. Chopping a long video into independent short clips introduces flickering at the seams. ATGS abandons per-Gaussian trajectory tracking entirely.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
