Research map / Architectures and efficiency / Attention and transformers
HeSS: Head Sensitivity Score for Sparsity Redistribution in VGGT
Yongsung Kim, Wooseok Song, Jaihyun Lew, Hun Hwangbo, Jaehoon Lee, Sungroh Yoon
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 |
|---|---|---|---|
| redistribute attention budget across heads by HeSS instead of uniform sparsity | SparseVGGT on VGGT · CO3Dv2 and DTU | AUC@30, ATE, Chamfer | Improved |
| HeSS-guided sparsification overhead | SparseVGGT · CO3Dv2 runtime | runtime (s): 8.42 -> 8.37 | No clear effect |
| add iterative budget capping and reallocation (ablation) | HeSS-Guided Sparsification · CO3Dv2 and DTU | AUC@30, Chamfer | Improved |
| use W_Q instead of W_K for Fisher computation (ablation) | HeSS · DTU | Chamfer distance: 1.9167 -> 1.6030 | Improved |
| use W_Q instead of W_V for Fisher computation (ablation) | HeSS · DTU | Chamfer distance: 1.9661 -> 1.6030 | Improved |
| sum-normalization instead of linear, softmax or log scaling (ablation) | HeSS combination · DTU | Chamfer distance: 1.8398/1.9685/2.1413 -> 1.6030 | Improved |
| combine HeSS_cam and HeSS_pc rather than using only one (ablation) | HeSS · DTU | Chamfer distance | Improved |
| invert HeSS ranking (prune sensitive heads first) (ablation) | HeSS-Guided Sparsification · CO3Dv2 and DTU | AUC@30, Chamfer | Hurt |
| apply HeSS-guided sparsification to pi3 with lambda=0 | pi3 · CO3Dv2 and DTU | AUC@30, Chamfer | Improved |
| replace head-level ViT pruning baselines with 3D-aware HeSS sparsification | VGGT · CO3Dv2 | AUC@30 | Improved |
Closest papers by meaning
- RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer (NeurIPS 2026 · similarity 0.88)
- QVGGT: Post-Training Quantized Visual Geometry Grounded Transformer (CVPR 2026 · similarity 0.88)
- TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention (NeurIPS 2026 · similarity 0.87)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.87)
- DSVT: Dynamic Sparse Voxel Transformer With Rotated Sets (CVPR 2023 · similarity 0.87)
- SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer (CVPR 2023 · 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.
