Research map / Detection and segmentation / Weakly supervised detection and segmentation
PreDet: Large-Scale Weakly Supervised Pre-Training for Detection
Vignesh Ramanathan, Rui Wang, Dhruv Mahajan
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 detection-specific query-box lookup losses to IG-50M classification pre-training | Mask R-CNN ResNeXt-101-32x8d · MS-COCO 1x | APbox: 44.4 -> 47.1 | Improved |
| add classification loss Lcls to the detection losses (ablation) | PreDet · MS-COCO 1x | APbox: 43.7 -> 47.1 | Improved |
| add CROI-head losses to the classification-only model (ablation) | Lcls model · MS-COCO 1x | APbox: 44.6 -> 46.3 | Improved |
| add CROI regression loss (ablation) | PreDet CROI-head · MS-COCO 1x | APbox: 45.9 -> 46.3 | Improved |
| add CRPN losses (ablation) | PreDet · MS-COCO 1x | APbox: 46.3 -> 47.1 | Improved |
| pre-train on IG-50M instead of ImageNet | PreDet · MS-COCO 1x | APbox: 45.8 -> 47.1 | Improved |
| PreDet-IG50M vs InfoMin on localization | Mask R-CNN · MS-COCO | APbox75: 49.9 -> 51.6 | Improved |
| PreDet pre-training vs training from scratch at full schedule | Mask R-CNN · MS-COCO | APbox: 45.8 -> 47.1 | Improved |
| use PreDet-IG50M instead of best baseline on RetinaNet | RetinaNet · MS-COCO | APbox: 43.0 -> 45.1 | Improved |
| PreDet-IG50M instead of SEER on PASCAL VOC | Faster R-CNN · PASCAL VOC07 test | APbox: 58.6 -> 62.5 | Improved |
| PreDet-IG50M instead of ImageNet pre-training with small target set | Mask R-CNN · MS-COCO 1k images | APbox: 11.3 -> 16.1 | Improved |
| larger backbone capacity (101-32x16d vs 50-32x4d) (ablation) | PreDet vs ImageNet · MS-COCO 1x | APbox gain: 2.3% -> 3.6% | Improved |
Closest papers by meaning
- ELDET: Early-Learning Distillation with Noisy Labels for Object Detection (NeurIPS 2025 · similarity 0.91)
- AlignDet: Aligning Pre-training and Fine-tuning in Object Detection (ICCV 2023 · similarity 0.91)
- DAP: Detection-Aware Pre-Training With Weak Supervision (CVPR 2021 · similarity 0.90)
- Instance Localization for Self-Supervised Detection Pretraining (CVPR 2021 · similarity 0.90)
- Rethinking ImageNet Pre-Training (ICCV 2019 · similarity 0.90)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · 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.
