Research map / Detection and segmentation / CNN features
Conditional DETR for Fast Training Convergence
Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang
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 conditional cross-attention (conditional spatial query) | DETR-R50, 50 epochs · COCO 2017 val (R50, 50 epochs) | AP: 34.9 -> 40.9 | Improved |
| add conditional cross-attention at 50 epochs | DETR-DC5-R50 · COCO 2017 val (DC5-R50, 50 epochs) | AP: 36.7 -> 43.8 | Improved |
| add conditional cross-attention at 50 epochs | DETR-R101 · COCO 2017 val (R101, 50 epochs) | AP: 36.9 -> 42.8 | Improved |
| add conditional cross-attention at 50 epochs | DETR-DC5-R101 · COCO 2017 val (DC5-R101, 50 epochs) | AP: 38.6 -> 45.0 | Improved |
| train 108 epochs instead of 50 | Conditional DETR-R50 · COCO 2017 val (R50, 50 epochs) | AP: 40.9 -> 43.0 | Improved |
| add learned or predicted reference points instead of s=(0,0) (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 36.8 -> 40.9 | Improved |
| learn reference points as parameters instead of predicting from object query (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 40.7 -> 40.9 | No clear effect |
| use full conditional spatial query instead of positional embedding only (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 37.8 -> 40.9 | Improved |
| use full conditional spatial query instead of transformation only (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 37.6 -> 40.9 | Improved |
| use transformation from decoder embedding f rather than content embedding in the query (CSQ-C) (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 37.1 -> 40.9 | Improved |
| predict transformation from decoder embedding instead of self-attention output (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 40.2 -> 40.9 | Improved |
| add conditional spatial query on top of focal loss and offset regression (ablation) | Conditional DETR · COCO 2017 val (R50, 50 epochs) | AP: 36.9 -> 40.9 | Improved |
Closest papers by meaning
- Decoupled DETR: Spatially Disentangling Localization and Classification for Improved End-to-End Object Detection (ICCV 2023 · similarity 0.90)
- DETR With Additional Global Aggregation for Cross-Domain Weakly Supervised Object Detection (CVPR 2023 · similarity 0.89)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.89)
- Unsupervised Pre-Training for Detection Transformers (TPAMI 2022 · similarity 0.89)
- Dynamic DETR: End-to-End Object Detection With Dynamic Attention (ICCV 2021 · similarity 0.89)
- FP-DETR: Detection Transformer Advanced by Fully Pre-training (ICLR 2022 · 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.
