Research map / Detection and segmentation / CNN features
Rethinking Transformer-Based Set Prediction for Object Detection
Zhiqing Sun, Shengcao Cao, Yiming Yang, Kris M. Kitani
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 set prediction loss (ablation) | TSP-RCNN (ResNet-50) · COCO val | AP: 42.7 -> 43.8 | Improved |
| add set prediction loss (ablation) | TSP-RCNN (ResNet-101) · COCO val | AP: 44.0 -> 44.8 | Improved |
| add positional encoding (ablation) | TSP-RCNN (ResNet-50) · COCO val | AP: 43.4 -> 43.8 | Improved |
| add positional encoding (ablation) | TSP-RCNN (ResNet-101) · COCO val | AP: 44.4 -> 44.8 | Improved |
| add Transformer encoder set prediction (TSP-FCOS) to FCOS | FCOS ResNet-50 · COCO val | AP: 41.0 -> 43.1 | Improved |
| add Transformer encoder set prediction (TSP-RCNN) to Faster RCNN | Faster RCNN-FPN ResNet-50 · COCO val | AP: 40.2 -> 43.8 | Improved |
| enlarge baseline to match FLOPs instead of adding Transformer | Faster RCNN ResNet-50 · COCO val | AP: 40.2 -> 40.9 | No clear effect |
| enlarge baseline to match FLOPs instead of adding Transformer | FCOS ResNet-50 · COCO val | AP: 41.0 -> 41.5 | No clear effect |
| longer 96-epoch schedule with random crop (TSP-RCNN+) | TSP-RCNN ResNet-50 · COCO val | AP: 43.8 -> 45.0 | Improved |
| increase number of proposals from 100 to 700 (ablation) | TSP-RCNN ResNet-50 · COCO val | AP: 40.3 -> 43.8 | Improved |
| increase feature positions from 100 to 700 (ablation) | TSP-FCOS ResNet-50 · COCO val | AP: 40.0 -> 43.1 | Improved |
| increase proposals from 500 to 700 (ablation) | TSP-RCNN ResNet-50 · COCO val | AP: 43.7 -> 43.8 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- UP-DETR — UP-DETR: Unsupervised Pre-Training for Object Detection With Transformers (CVPR 2021) · better on 1 of 1 shared result
- Dynamic RCNN — Dynamic R-CNN: Towards High Quality Object Detection via Dynamic Training (ECCV 2020) · better on 1 of 1 shared result
- FCOS — FCOS: Fully Convolutional One-Stage Object Detection (ICCV 2019) · better on 1 of 1 shared result
- Faster-RCNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · 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
- End-to-End Object Detection with Transformers (ECCV 2020 · similarity 0.93)
- Anchor DETR: Query Design for Transformer-Based Detector (AAAI 2022 · similarity 0.89)
- Efficient Decoder-Free Object Detection with Transformers (ECCV 2022 · similarity 0.88)
- Mr. DETR: Instructive Multi-Route Training for Detection Transformers (CVPR 2025 · similarity 0.87)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.87)
- KD-DETR: Knowledge Distillation for Detection Transformer with Consistent Distillation Points Sampling (CVPR 2024 · similarity 0.87)
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.
