Research map / Detection and segmentation / CNN features
Deeply Unsupervised Patch Re-Identification for Pre-Training Object Detectors
Jian Ding, Enze Xie, Hang Xu, Chenhan Jiang, Zhenguo Li, Ping Luo, Gui-Song Xia
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 patch-level re-identification loss on conv5_x (beta 0,0,0,1) (ablation) | DUPR · PASCAL VOC detection | AP: 57.6 -> 58.3 | Improved |
| add patch loss on conv5_x (beta 0,0,0,1) (ablation) | DUPR · COCO detection (R50-FPN) | box AP: 39.2 -> 39.5 | Improved |
| add patch loss on conv4_x and conv5_x (beta 0,0,1,1) (ablation) | DUPR · COCO detection (R50-FPN) | box AP: 39.5 -> 40.0 | Improved |
| add multi-level image contrastive loss alpha=(0.1,0.4,0.7,1) to final-layer only (ablation) | MoCo v2 · PASCAL VOC detection | AP: 57.0 -> 57.6 | Improved |
| use equal weights alpha=(1,1,1,1) on all layers (ablation) | MoCo v2 · COCO Mask R-CNN | box AP: 38.9 -> 38.5 | Hurt |
| add intermediate-layer loss alpha=(0.1,0.4,0.7,1) (ablation) | MoCo v2 · VOC SVM classification conv3_x | accuracy: 58.8 -> 68.8 | Improved |
| add intermediate-layer loss alpha=(0.1,0.4,0.7,1) (ablation) | MoCo v2 · VOC SVM classification conv5_x | accuracy: 84.1 -> 83.7 | Hurt |
| increase RoI size from 1 to 7 for patch loss (ablation) | DUPR · COCO detection | box AP: 39.1 -> 39.5 | Improved |
| increase RoI size beyond 7 to 11 for patch loss (ablation) | DUPR · VOC detection | AP: 58.3 -> 58.4 | No clear effect |
| replace image-level with patch-level contrastive loss at conv5_x (ablation) | MoCo v2 · VOC detection | AP75: 63.6 -> 64.6 | Improved |
| replace image-level with patch-level contrastive loss at conv5_x (ablation) | MoCo v2 · COCO detection | box AP: 38.9 -> 38.8 | No clear effect |
| use DUPR pre-training instead of MoCo v2 with few fine-tune iterations (12k) | Mask R-CNN R50-FPN · COCO | mAP: +2.9 mAP | Improved |
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.
- DetCo — DetCo: Unsupervised Contrastive Learning for Object Detection (ICCV 2021) · better on 0 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
- Patch-Based Discriminative Feature Learning for Unsupervised Person Re-Identification (CVPR 2019 · similarity 0.89)
- Unsupervised Object Detection with Theoretical Guarantees (NeurIPS 2024 · similarity 0.88)
- Unsupervised Object Detection Pretraining with Joint Object Priors Generation and Detector Learning (NeurIPS 2022 · similarity 0.88)
- DETReg: Unsupervised Pretraining With Region Priors for Object Detection (CVPR 2022 · similarity 0.88)
- No Pains, More Gains: Recycling Sub-Salient Patches for Efficient High-Resolution Image Recognition (CVPR 2025 · similarity 0.86)
- Catalyst for Clustering-Based Unsupervised Object Re-identification: Feature Calibration (AAAI 2024 · similarity 0.85)
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.
