Research map / Architectures and efficiency / Attention and transformers
Enhancing Vision Transformers for Object Detection via Context-Aware Token Selection and Packing
Tianyi Zhang, Baoxin Li, Jae-sun Seo, Yu Cao
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 |
|---|---|---|---|
| add SPA dynamic token selection instead of uniform top-50 (ablation) | SPT-T · PASCAL VOC 2012 | mAP: 44.42 -> 44.49 | Improved |
| add select loss Lselect (ablation) | SPT-T · PASCAL VOC 2012 | mAP: 44.49 -> 44.60 | Improved |
| add select loss Lselect (lowers mean select ratio) (ablation) | SPT-T · PASCAL VOC 2012 | mean select ratio: 59.77 -> 29.60 | Improved |
| start SPA at the third stage rather than only the 4th stage (ablation) | SPT-T · BDD100K | AP: 21.9 -> 22.6 | Improved |
| start SPA at the third stage rather than the 2nd stage (ablation) | SPT-T · BDD100K | AP: 20.5 -> 22.6 | Improved |
| start SPA at the third stage rather than in all four stages (ablation) | SPT-T · BDD100K | AP: 18.3 -> 22.6 | Improved |
| weight positive tokens 5x in select loss (ablation) | SPT · BDD-S | mAP: 33% -> 71% select ratio, mAP dropped to 5.3 | Hurt |
| use SPT-T/16 (more blocks) rather than SPT-T/12 (ablation) | SPT-T · BDD100K | AP: 22.3 -> 22.6 | Improved |
| replace DynamicSwin-T backbone with SPT-T | Cascade Mask RCNN · COCO2017 | AP: 44.3 -> 47.1 | Improved |
| replace DynamicSwin-S backbone with SPT-S | Cascade Mask RCNN · COCO2017 | AP: 47.2 -> 49.3 | Improved |
| replace DynamicSwin-B backbone with SPT-B | Cascade Mask RCNN · COCO2017 | AP: 50.5 -> 53.2 | Improved |
| replace DynamicSwin-S backbone with SPT-S for instance segmentation | Mask RCNN · COCO2017 instance segmentation | AP: 39.6 -> 40.9 | Improved |
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.
- SparseViT — SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer (CVPR 2023) · better on 1 of 1 shared result
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
- Token Selection is a Simple Booster for Vision Transformers (TPAMI 2022 · similarity 0.89)
- DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification (NeurIPS 2021 · similarity 0.89)
- Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers (ICML 2025 · similarity 0.88)
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.88)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.87)
- PADRe: A Unifying Polynomial Attention Drop-in Replacement for Efficient Vision Transformer (ICLR 2025 · similarity 0.86)
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.
