Research map / Detection and segmentation / CNN features
Proposal-Contrastive Pretraining for Object Detection from Fewer Data
Read the paperCheck 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 |
|---|---|---|---|
| Localization-aware positives in contrastive loss (LocSCE, IoU threshold 0.5) vs original SCE (ablation) | ProSeCo · Mini-COCO 5% | mAP: +1.7 | Helped |
| Use a fixed backbone pretrained with local-info (SCRL) instead of global-info (SwAV) (ablation) | ProSeCo · Mini-COCO 5% | mAP: +1 p.p. | Helped |
| Batch size 192 instead of 64 (ablation) | ProSeCo · Mini-COCO 5% | mAP: -0.2 | Hurt |
| Large-scale images in weak augmentation vs mid-scale (ablation) | ProSeCo · Mini-COCO 5% | mAP: +0.7 | Helped |
| Temperature tau_t 0.07 vs 0.05 (ablation) | ProSeCo · Mini-COCO 5% | mAP: +0.3 | Helped |
| EMA keep rate 0.999 vs 0.99/0.996 (ablation) | ProSeCo · Mini-COCO 5% | mAP: +0.1 over 0.996 is 1.0 difference; text reports +0.1 p.p. overall | Helped |
| InfoNCE instead of SCE (no localization) (ablation) | ProSeCo · Mini-COCO 5% | mAP: +0.3 | Helped |
| Localization positives on InfoNCE (LocNCE) vs SCE-based LocSCE (ablation) | ProSeCo · Mini-COCO 5% | mAP: LocSCE +0.8 over LocNCE | Mixed |
| ProSeCo pretraining vs DETReg on Def. DETR across Mini-COCO, FSOD, and VOC settings | Def. DETR · Mini-COCO 1%/5%/10%, FSOD-test, Mini-VOC, PASCAL VOC | mAP: +2.2 to +4.1 over DETReg | Helped |
| ProSeCo pretraining on full COCO 1x finetuning | Def. DETR · COCO full (118k) | mAP: 0.0 vs DETReg | No clear effect |
| Pretraining on COCO instead of ImageNet (ablation) | ProSeCo · Mini-COCO 5% | mAP: -0.4 | Hurt |
Closest papers by meaning
- Aligning Pretraining for Detection via Object-Level Contrastive Learning (NeurIPS 2021 · similarity 0.92)
- Point-Level Region Contrast for Object Detection Pre-Training (CVPR 2022 · similarity 0.90)
- FSCE: Few-Shot Object Detection via Contrastive Proposal Encoding (CVPR 2021 · similarity 0.90)
- Unsupervised Object Detection Pretraining with Joint Object Priors Generation and Detector Learning (NeurIPS 2022 · similarity 0.88)
- Contrastive Mean Teacher for Domain Adaptive Object Detectors (CVPR 2023 · similarity 0.88)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 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.