Research map / 3D vision / Depth and stereo
AVGGT: Rethinking Global Attention for Accelerating VGGT
Xianbing Sun, Zhikai Zhu, Zhengyu Lou, Bo Yang, Jinyang Tang, Liqing Zhang, He Wang, Jianfu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| convert early global attention layers to frame attention (ablation) | VGGT · RealEstate10K (10 frames) | AUC@5: 63.176 -> 62.826 | No clear effect |
| replace early global K/V with a single mean token (ablation) | VGGT · RealEstate10K (10 frames) | AUC@5: 63.176 -> 61.788 | No clear effect |
| keep last global attention layers (not converted to frame attention) (ablation) | AVGGT(2) · RealEstate10K (10 frames) | AUC@5: 59.637 -> 61.959 | Improved |
| subsample queries too (nearest query) (ablation) | AVGGT(2) · RealEstate10K (10 frames) | AUC@5: 59.583 -> 42.565 | Hurt |
| subsample queries too (global-mean query) (ablation) | AVGGT(2) · RealEstate10K (10 frames) | AUC@5: 59.583 -> 26.973 | Hurt |
| select K/V tokens on a fixed grid instead of SIFT keypoints (ablation) | AVGGT(2) · RealEstate10K (10 frames) | AUC@5: 55.438 -> 61.959 | Improved |
| select K/V tokens on a fixed grid instead of randomly (ablation) | AVGGT(2) · RealEstate10K (10 frames) | AUC@5: 57.849 -> 61.959 | Improved |
| add diagonal preservation and mean component to subsampling (ablation) | AVGGT(2) · 7-Scenes (333 frames) | AUC@5: 24.539 -> 26.061 | Improved |
| add diagonal preservation and mean component to subsampling (ablation) | AVGGT(2) · RealEstate10K (10 frames) | AUC@5: 61.908 -> 61.959 | No clear effect |
| keep first-frame tokens uncompressed (ablation) | AVGGT(4) · RealEstate10K (10 frames) | AUC@5: 55.570 -> 59.583 | Improved |
| increase subsampling factor from 2 to 9 (ablation) | AVGGT · RealEstate10K (10 frames) | AUC@5: 61.959 -> 53.257 | Hurt |
| increase subsampling factor from 2 to 9 (ablation) | A_pi3 · RealEstate10K (10 frames) | AUC@5: 64.832 -> 40.437 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 7 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention (NeurIPS 2026 · similarity 0.89)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.88)
- MVVBench: Benchmarking 4D Reasoning in Vision-Language Models (NeurIPS 2026 · similarity 0.87)
- The 3D-PC: a benchmark for visual perspective taking in humans and machines (ICLR 2025 · similarity 0.87)
- RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer (NeurIPS 2026 · similarity 0.87)
- CanViT: Toward Active-Vision Foundation Models (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
