Research map / 3D vision / Depth and stereo
Block-Sparse Global Attention for Efficient Multi-View Geometry Transformers
Chung-Shien Brian Wang, Christian Schmidt, Jens Piekenbrinck, Bastian Leibe
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 |
|---|---|---|---|
| replace dense global attention with training-free block-sparse attention | VGGT, π3, MapAnything · multi-view pose and pointmap benchmarks | task accuracy vs inference speed: more than 3× faster, comparable accuracy | Improved |
| increase sparsity ratio of global attention (ablation) | Sparse VGGT / π3 · relative pose estimation (RE10K, CO3Dv2, TUM, ScanNet) | pose accuracy | Hurt |
| compute dense attention for special (camera/register) tokens (ablation) | Block-sparse attention · CO3Dv2 / ETH3D | AUC@30 / Chamfer | Improved |
| combine top-k with CDF thresholding instead of CDF-only (SpargeAttention) (ablation) | Block-sparse attention · CO3Dv2 / ETH3D | AUC@30 / Chamfer | Improved |
| random block masking instead of predicted block masks (ablation) | Block-sparse attention · CO3Dv2 / ETH3D | AUC@30 / Chamfer | Hurt |
| add linear projection after pooling queries and keys (ablation) | Block mask predictor · VGGT | task performance | No clear effect |
| skip a single middle aggregator global-attention layer (ablation) | VGGT · CO3Dv2 | AUC@30 | Hurt |
Closest papers by meaning
- AVGGT: Rethinking Global Attention for Accelerating VGGT (CVPR 2026 · similarity 0.91)
- TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention (NeurIPS 2026 · similarity 0.90)
- XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025 · similarity 0.89)
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.88)
- GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers (ICLR 2024 · similarity 0.88)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · 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.
