Research map / Detection and segmentation / CNN features
Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace RoIPooling with RoIAlign (ablation) | Swin-T Faster R-CNN trained from scratch · MS COCO | mAP: 26.6 -> 30.3 | Helped |
| Replace RoIPooling with RoIAlign (ablation) | Swin-T Faster R-CNN with ImageNet pre-training · MS COCO | mAP: 42.1 -> 42.5 | Helped |
| Replace the first two transformer stages with convolution stages (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 30.3 -> 37.9 | Helped |
| Replace the first two transformer stages with convolution stages (ablation) | FCOS trained from scratch · MS COCO | mAP: 23.6 -> 29.5 | Helped |
| Use convolution in all four backbone stages instead of C-C-T-T (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 45.8 -> 42.3 | Hurt |
| Use convolution in the first three backbone stages instead of C-C-T-T (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 45.8 -> 43.8 | Hurt |
| Use convolution only in the first backbone stage instead of C-C-T-T (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 45.8 -> 44.8 | Hurt |
| Use C-C-T-T instead of the fully transformer T-T-T-T backbone (ablation) | Faster R-CNN with ImageNet pre-training · MS COCO | mAP: 45.5 -> 45.3 | No clear effect |
| Increase backbone scale from Swin-T to Swin-S while using C-C-T-T (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 45.8 -> 48.6 | Helped |
| Increase backbone scale from Swin-S to Swin-B while using C-C-T-T (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 48.6 -> 51.2 | Helped |
| Increase backbone scale from Swin-B to Swin-L while using C-C-T-T (ablation) | Faster R-CNN trained from scratch · MS COCO | mAP: 51.2 -> 53.0 | Helped |
| Add gradient calibration to the C-C-T-T model (ablation) | Faster R-CNN and FCOS · MS COCO | mAP | Helped |
Closest papers by meaning
- An Empirical Study of Training End-to-End Vision-and-Language Transformers (CVPR 2022 · similarity 0.89)
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021 · similarity 0.89)
- FP-DETR: Detection Transformer Advanced by Fully Pre-training (ICLR 2022 · similarity 0.89)
- ViDT: An Efficient and Effective Fully Transformer-based Object Detector (ICLR 2022 · similarity 0.89)
- ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision (ICML 2021 · similarity 0.88)
- Vision Transformers Need Registers (ICLR 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.