Research map / Architectures and efficiency / Attention and transformers
SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use window-level activation pruning instead of token-level pruning | Swin-T on nuScenes · nuScenes | latency: 23.8 vs 32.6 | Helped |
| Prune more from a higher-resolution input (ablation) | SparseViT on nuScenes · nuScenes | mAP: 28.7 -> 30.0 | Helped |
| Use non-uniform layerwise sparsity instead of uniform pruning (ablation) | SparseViT on nuScenes · nuScenes | mAP: +0.6% | Helped |
| Use evolutionary search instead of random search (ablation) | SparseViT on COCO · COCO | validation mAP: 37.5 | Helped |
| Use sparsity-aware adaptation to estimate accuracy before finetuning (ablation) | SparseViT on nuScenes · nuScenes | mAP: 28.31 vs 30.93 (without SAA); 30.83 vs 30.63 (with SAA) | Helped |
| Add sparsity-aware adaptation during finetuning (ablation) | SparseViT on nuScenes · nuScenes | mAP: 30.83 -> 31.21 | Helped |
| Use L2 magnitude scoring instead of learned window scoring (ablation) | SparseViT on nuScenes · nuScenes | mAP drop: -0.6 -> 0.0 | Helped |
| Share window scores per stage instead of scoring independently per block (ablation) | SparseViT on nuScenes · nuScenes | latency: 23.3 -> 23.0 | Helped |
| Execute FFN and layer normalization window-wise (ablation) | Swin Transformer · nuScenes | accuracy: less than 0.1% | No clear effect |
| Duplicate unselected-window features rather than discard them (ablation) | SparseViT · Dense prediction tasks | accuracy | Helped |
Closest papers by meaning
- Chasing Sparsity in Vision Transformers: An End-to-End Exploration (NeurIPS 2021 · similarity 0.92)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.89)
- LookThere! Sparse Vision by Reinforced Selection (NeurIPS 2026 · similarity 0.88)
- Bootstrapping SparseFormers from Vision Foundation Models (CVPR 2024 · similarity 0.88)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.88)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.88)
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.