Research map / Detection and segmentation / CNN features
End-to-End Object Detection with Transformers
Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko
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 transformer encoder layers (0 to 6) (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 36.7 -> 40.6 | Improved |
| add 12 encoder layers instead of 6 (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 40.6 -> 41.6 | Improved |
| add FFN in transformer layers (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: drops 2.3 AP when removed (41.3M to 28.7M params) | Improved |
| add spatial positional encodings (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 32.8 -> 40.6 | Improved |
| pass sine positional encodings at every attention layer instead of once at input (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 39.2 -> 40.6 | Improved |
| use spatial positional encodings in the encoder (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 39.3 -> 40.6 | Improved |
| add GIoU loss to class+L1 (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 35.8 -> 40.6 | Improved |
| add L1 loss to class+GIoU (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: 39.9 -> 40.6 | Improved |
| apply NMS to later decoder layer outputs (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP | Hurt |
| add decoder layers (first to last layer outputs) (ablation) | DETR R50 · COCO val 2017 (bbox, 300-epoch ablation schedule) | AP: +8.2/9.5 AP between first and last layer | Improved |
| add random crop augmentation | DETR · COCO val 2017 (bbox) | AP: approximately 1 AP | Improved |
| longer schedule (500 vs 300 epochs) | DETR · COCO val 2017 (bbox) | AP: 40.6 -> 42.0 | 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.
- UPSnet — UPSNet: A Unified Panoptic Segmentation Network (CVPR 2019) · 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
- Efficient Decoder-Free Object Detection with Transformers (ECCV 2022 · similarity 0.94)
- Rethinking Transformer-Based Set Prediction for Object Detection (ICCV 2021 · similarity 0.93)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.93)
- WB-DETR: Transformer-Based Detector Without Backbone (ICCV 2021 · similarity 0.91)
- UP-DETR: Unsupervised Pre-Training for Object Detection With Transformers (CVPR 2021 · similarity 0.91)
- CF-DETR: Coarse-to-Fine Transformers for End-to-End Object Detection (AAAI 2022 · similarity 0.91)
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.
