Research map / Detection and segmentation / CNN features
Unsupervised Pre-Training for Detection Transformers
Zhigang Dai, Bolun Cai, Yugeng Lin, Junying Chen
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 |
|---|---|---|---|
| Unsupervised transformer pre-training (150 epochs) | UP-DETR · PASCAL VOC | AP: 49.9 -> 56.1 | Improved |
| Unsupervised transformer pre-training (300 epochs) | UP-DETR · PASCAL VOC | AP: 54.1 -> 57.2 | Improved |
| Unsupervised pre-training (150 epochs) | UP-DETR · COCO val2017 | AP: 39.7 -> 40.5 | Improved |
| Unsupervised pre-training (300 epochs) | UP-DETR · COCO val2017 | AP: 42.1 -> 43.1 | Improved |
| Unsupervised pre-training for one-shot detection | UP-DETR · VOC07 one-shot seen classes | AP50: 46.2 -> 69.0 | Improved |
| Unsupervised pre-training for one-shot detection | UP-DETR · VOC07 one-shot unseen classes | AP50: 57.3 -> 73.1 | Improved |
| Unsupervised pre-training for panoptic segmentation | UP-DETR · COCO panoptic val2017 | PQ: 44.3 -> 44.7 | Improved |
| Use multi-query patches M=10 instead of single-query M=1 (ablation) | UP-DETR · PASCAL VOC (ablation) | AP: 53.1 -> 54.9 | Improved |
| Freeze the CNN backbone during pre-training (ablation) | UP-DETR · PASCAL VOC (ablation) | AP50: 74.0 -> 78.7 | Improved |
| Add feature reconstruction with frozen CNN (ablation) | UP-DETR · PASCAL VOC (ablation) | AP50: 78.7 -> 78.7 | No clear effect |
| Add feature reconstruction without frozen CNN (ablation) | UP-DETR · PASCAL VOC (ablation) | AP50: 74.0 -> 62.0 | Hurt |
| Add attention mask in pre-training (ablation) | UP-DETR · PASCAL VOC fine-tuning | AP | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Faster R-CNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 1 of 3 shared results
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
- UP-DETR: Unsupervised Pre-Training for Object Detection With Transformers (CVPR 2021 · similarity 0.98)
- FP-DETR: Detection Transformer Advanced by Fully Pre-training (ICLR 2022 · similarity 0.91)
- End-to-End Object Detection with Transformers (ECCV 2020 · similarity 0.90)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.90)
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 2024 · similarity 0.88)
- Salience DETR: Enhancing Detection Transformer with Hierarchical Salience Filtering Refinement (CVPR 2024 · similarity 0.88)
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.
