Research map / Detection and segmentation / CNN features
AugDETR: Improving Multi-scale Learning for Detection Transformer
Jinpeng Dong, Yutong Lin, Chen Li, Sanping Zhou, Nanning Zheng
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 Hybrid Attention Encoder and Encoder-Mixing Cross-Attention to DINO | DINO R50 12 epochs · COCO val2017 | AP: 49.0 -> 50.2 | Improved |
| add AugDETR to DDQ | DDQ · COCO val2017 | AP: 50.8 -> 51.8 | Improved |
| add AugDETR to DINO with Swin-T | DINO Swin-T · COCO val2017 | AP: 51.3 -> 52.3 | Improved |
| add AugDETR to DINO under 2x schedule | DINO 24 epochs · COCO val2017 | AP: 50.4 -> 51.3 | Improved |
| add Hybrid Attention Encoder alone (ablation) | DINO · COCO val2017 | AP: 49.0 -> 49.6 | Improved |
| add Encoder-Mixing Cross-Attention alone (ablation) | DINO · COCO val2017 | AP: 49.0 -> 49.8 | Improved |
| increase deformable encoder sampling points from 4 to 6 or 8 (ablation) | DINO deformable encoder · COCO val2017 | AP: 49.0 -> 48.4 (6 points), 48.5 (8 points) | Hurt |
| use hybrid attention in all 6 encoder layers (ablation) | HAE · COCO val2017 | AP: 49.0 -> 48.8 | No clear effect |
| stack more encoder layers (10) without EMCA (ablation) | DINO · COCO val2017 | AP: 49.0 -> 47.8 | Hurt |
| add iterative encoder cross-attention (ablation) | DINO · COCO val2017 | AP: 49.0 -> 47.7 | Hurt |
| use max fusion instead of sigmoid weighted fusion in EMCA (ablation) | EMCA · COCO val2017 | AP: 49.0 -> 48.6 | Hurt |
| use sigmoid weighted fusion rather than sum fusion (ablation) | EMCA · COCO val2017 | AP: 49.3 -> 49.8 | Improved |
Closest papers by meaning
- Lite DETR: An Interleaved Multi-Scale Encoder for Efficient DETR (CVPR 2023 · similarity 0.90)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.90)
- DESTR: Object Detection With Split Transformer (CVPR 2022 · similarity 0.89)
- Dynamic DETR: End-to-End Object Detection With Dynamic Attention (ICCV 2021 · similarity 0.89)
- DAC-DETR: Divide the Attention Layers and Conquer (NeurIPS 2023 · similarity 0.88)
- Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement (CVPR 2024 · 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.
