Research map / Detection and segmentation / CNN features
Integrally Migrating Pre-trained Transformer Encoder-decoders for Visual Object Detection
Feng Liu, Xiaosong Zhang, Zhiliang Peng, Zonghao Guo, Fang Wan, Xiangyang Ji, Qixiang Ye
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 |
|---|---|---|---|
| replace randomly initialized conv head with randomly initialized transformer decoder (ablation) | Faster R-CNN ViT-S · MS COCO val (1x) | AP: 42.4 -> 42.2 | Hurt |
| pre-train the migrated decoder with MAE (ablation) | Faster R-CNN ViT-S with decoder head · MS COCO val (1x) | AP: 42.2 -> 42.5 | Improved |
| remove FPN from the feature extraction path (ablation) | imTED ViT-S · MS COCO val (1x) | AP: 42.5 -> 43.2 | Improved |
| add multi-scale feature modulator (MFM) (ablation) | imTED ViT-S · MS COCO val (1x) | AP: 43.2 -> 44.0 | Improved |
| increase decoder depth from 1 to 4 blocks (ablation) | imTED ViT-S · MS COCO val (1x) | AP: 42.3 -> 44.0 | Improved |
| increase decoder depth from 3 to 4 blocks (ablation) | imTED ViT-S · MS COCO val (1x) | AP: 43.9 -> 44.0 | No clear effect |
| integral migration of encoder-decoder with MFM versus MAE ViT-B baseline detector | Faster R-CNN ViT-B · MS COCO (3x) | AP: 50.5 -> 52.9 | Improved |
| use pre-trained decoder in single-object localization feature extractor (ablation) | MAE ViT encoder · ImageNet Localization | mAP: 43.4 -> 44.8 | Improved |
| use imTED instead of baseline detector for few-shot detection | Faster R-CNN ViT-B · MS COCO 30-shot | AP: 22.2 -> 30.2 | Improved |
| apply imTED components on occluded-object subset (ablation) | Faster R-CNN ViT-S · MS COCO occluded subset | AP: 36.6 -> 39.0 | Improved |
| imTED (ViT-L, 36 epochs) versus ViTDet on AP mask | Mask R-CNN ViT-L · MS COCO | AP mask: 48.9 (ViTDet) vs 48.1 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Meta-DETR — Meta-DETR: Image-Level Few-Shot Detection with Inter-Class Correlation Exploitation (TPAMI 2022) · better on 1 of 1 shared result
- DeFRCN — DeFRCN: Decoupled Faster R-CNN for Few-Shot Object Detection (ICCV 2021) · 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 0.90)
- Unleashing Vanilla Vision Transformer with Masked Image Modeling for Object Detection (ICCV 2023 · similarity 0.90)
- Exploring Plain Vision Transformer Backbones for Object Detection (ECCV 2022 · similarity 0.89)
- Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer (CVPR 2022 · similarity 0.88)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · similarity 0.87)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 2025 · similarity 0.86)
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.
