RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer
RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer
This paper introduces RegimeVGGT, a training-free acceleration technique for the Visual Geometry Grounded Transformer (VGGT) that removes redundancy layer-by-layer while preserving spatial coverage and pose-critical attention paths.
The authors propose RegimeVGGT, a training-free compression method for Visual Geometry Grounded Transformer (VGGT) that achieves a 6.7x speedup by applying layer-wise spatially preserving redundancy removal based on spectral and causal analyses of functional regimes across network depths.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- code: not_reported
- weights: not_reported
- dataset: not_reported
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
eserves cross-frame spatial coverage and the pose-critical path through a phase-shifted spatial grid, a reference-frame anchor, and uncompressed camera/register tokens. Without any training, RegimeVGGT achieves a 6.7x speedup over the memory-optimized VGGT* at matched reconstruction quality. Project page: https://tracyou07.github.io/regimevggt/
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org