Research map / Detection and segmentation / Open-set and few-shot detection
Aligning Pretraining for Detection via Object-Level Contrastive Learning
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 |
|---|---|---|---|
| add selective-search object proposals for object-level contrastive learning (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 38.1 -> 40.6 | Helped |
| add FPN to pretraining without head (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 40.6 -> 40.2 | Hurt |
| add R-CNN head along with FPN to pretraining (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 40.2 -> 41.2 | Helped |
| add scale-aware assignment (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 41.2 -> 41.6 | Helped |
| add box jitter (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 41.6 -> 41.7 | No clear effect |
| add multi-view V3 (scale augmentation) (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 41.7 -> 42.3 | Helped |
| use selective search instead of random boxes (K=1) (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 41.4 -> 41.6 | Helped |
| vary proposals per image K from 4 (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 42.3 -> 41.6 (K=1), 41.2 (K=16) | Hurt |
| lower EMA momentum from 0.99 to 0.98 (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 42.3 -> 35.0 | Hurt |
| increase batch size from 2048 to 4096 (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 42.3 -> 41.4 | Hurt |
| vary V3 resolution (96 to 192) (ablation) | SoCo · COCO 1x R50-FPN | AP^bb: 42.0-42.3 | No clear effect |
| add extra view V4 (SoCo*) | SoCo 400ep · COCO 1x R50-FPN | AP^bb: 43.0 -> 43.2 | Helped |
Closest papers by meaning
- Proposal-Contrastive Pretraining for Object Detection from Fewer Data (ICLR 2023 · similarity 0.92)
- Point-Level Region Contrast for Object Detection Pre-Training (CVPR 2022 · similarity 0.90)
- Boost Supervised Pretraining for Visual Transfer Learning: Implications of Self-Supervised Contrastive Representation Learning (AAAI 2022 · similarity 0.89)
- Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.89)
- What Makes Instance Discrimination Good for Transfer Learning? (ICLR 2021 · similarity 0.88)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · 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.