Research map / Detection and segmentation / CNN features
Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity
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 |
|---|---|---|---|
| Update only 10% of encoder tokens using DAM saliency | Sparse DETR with Swin-T · COCO 2017 val2017 | AP: 48.0 -> 48.2 | Helped |
| Update only 10% of encoder tokens | Sparse DETR with Swin-T · COCO 2017 val2017 | FLOPs: 185G -> 113G | Helped |
| Update only 10% of encoder tokens | Sparse DETR with Swin-T · COCO 2017 val2017 | FPS: 15.4 -> 21.2 | Helped |
| Add encoder auxiliary loss to selected encoder tokens (ablation) | Sparse DETR with Swin-T and 10% token keeping · COCO 2017 val2017 | AP: 46.8 -> 48.2 | Helped |
| Add encoder auxiliary loss (ablation) | Sparse DETR with Swin-T and 12 encoder layers · COCO 2017 val2017 | training convergence | Helped |
| Use DAM rather than OS to select encoder tokens (ablation) | Sparse DETR with Swin-T and 10% token keeping · COCO 2017 val2017 | AP: 48.0 -> 48.2 | Helped |
| Use DAM rather than random token selection (ablation) | Sparse DETR with Swin-T and 10% token keeping · COCO 2017 val2017 | AP: 45.5 -> 48.2 | Helped |
| Replace BCE with ranking loss for DAM scoring (ablation) | Sparse DETR with Swin-T and 30% token keeping · COCO 2017 val2017 | AP: 49.1 -> 48.8 | Hurt |
| Replace BCE with smoothed L1 regression loss for DAM scoring (ablation) | Sparse DETR with Swin-T and 30% token keeping · COCO 2017 val2017 | AP: 49.1 -> 48.6 | Hurt |
| Increase encoder token keeping from 10% to 40% (ablation) | Sparse DETR with Swin-T · COCO 2017 val2017 | AP: 48.2 -> 49.2 | Helped |
| Increase encoder token keeping from 10% to 50% (ablation) | Sparse DETR with ResNet-50 · COCO 2017 val2017 | AP: 45.3 -> 46.3 | Helped |
| Use a Swin-B backbone instead of Swin-T (ablation) | Sparse DETR at 40% token keeping · COCO 2017 val2017 | AP: 49.2 -> 53.3 | Helped |
Closest papers by meaning
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.90)
- MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism (CVPR 2025 · similarity 0.90)
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.89)
- Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers (ICML 2025 · similarity 0.89)
- Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (CVPR 2024 · similarity 0.89)
- Sparse Semi-DETR: Sparse Learnable Queries for Semi-Supervised Object Detection (CVPR 2024 · similarity 0.89)
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.