Research map / Detection and segmentation / CNN features
WB-DETR: Transformer-Based Detector Without Backbone
Fanfan Liu, Haoran Wei, Wenzhe Zhao, Guozhen Li, Jingquan Peng, Zihao Li
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 LIE-T2T encoder layers (0 to 3 layers) (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 30.7 -> 38.5 -> 40.2 -> 40.3 | Improved |
| add LIE-T2T layers, 0 to 2, for small objects (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | APS: 10.5 -> 19.3 | Improved |
| increase LIE-T2T layers from 2 to 3 (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 40.2 -> 40.3 | No clear effect |
| use LIE-T2T instead of T2T (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 38.1 -> 40.2 | Improved |
| use LIE-T2T instead of T2T (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP75: 41.4 -> 44.8 | Improved |
| use LIE-T2T instead of T2T (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP50: 62.8 -> 63.2 | No clear effect |
| use smaller patch/step (8/4 vs 16/8) (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 40.2 -> 41.1 | Improved |
| use larger patch/step (32/16 vs 16/8) (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 40.2 -> 36.3 | Hurt |
| use non-overlapping patches (16/16 vs overlapping 16/8) (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 40.2 -> 35.4 | Hurt |
| use non-overlapping 8/8 vs overlapping 16/8 (ablation) | WB-DETR (2-8) · COCO val2017 (WB-DETR 2-8) | AP: 40.2 -> 39.7 | Hurt |
| remove CNN backbone: WB-DETR (2-12) vs DETR (ablation) | WB-DETR (2-12) · COCO val2017 | AP: DETR 42.0 vs WB-DETR (2-12) 41.8 | No clear effect |
| remove CNN backbone: WB-DETR (2-12) vs DETR | WB-DETR (2-12) · COCO val2017 | FPS: 28 -> 36 | Improved |
Closest papers by meaning
- Efficient Decoder-Free Object Detection with Transformers (ECCV 2022 · similarity 0.92)
- End-to-End Object Detection with Transformers (ECCV 2020 · similarity 0.91)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.90)
- KD-DETR: Knowledge Distillation for Detection Transformer with Consistent Distillation Points Sampling (CVPR 2024 · similarity 0.89)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.88)
- Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (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.
