Research map / Detection and segmentation / Open-set and few-shot detection
CYBORGS: Contrastively Bootstrapping Object Representations by Grounding in Segmentation
Renhao Wang, Hang Zhao, Yang Gao
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 periodic mask bootstrapping (N=100) instead of initial masks only (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 9.75 -> 58.0 | Improved |
| bootstrap every 100 epochs instead of every 10 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 52.6 -> 58.0 | Improved |
| bootstrap every 100 epochs instead of every 400 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 52.7 -> 58.0 | Improved |
| add CRF refinement of bootstrapped masks (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 58.0 -> 58.6 | No clear effect |
| use res2+res3+res4 features in contrastive objective instead of res2 only (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 55.1 -> 58.0 | Improved |
| use multi-layer features for KMeans (res2+3+4) instead of res2.b2 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 58.0 -> 55.0 | Hurt |
| use scale-dynamic K~U[2,256] instead of fixed K=81 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 42.5 -> 58.0 | Improved |
| use scale-dynamic K~U[2,256] instead of fixed K=2 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 22.6 -> 58.0 | Improved |
| use vMF clustering loss instead of Euclidean distance (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 53.8 -> 58.0 | Improved |
| add vMF clustering loss (M=5) versus none (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 41.8 -> 58.0 | Improved |
| apply vMF loss every epoch (M=1) instead of every 5 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 58.0 -> 58.6 | No clear effect |
| increase vMF loss weight from 0.001 to 0.1 (ablation) | CYBORGS · VOC detection ablation (COCO-pretrained) | AP: 42.3 -> 58.0 | Improved |
Closest papers by meaning
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.88)
- ContrastMask: Contrastive Learning To Segment Every Thing (CVPR 2022 · similarity 0.88)
- Bootstrapping Semantic Segmentation with Regional Contrast (ICLR 2022 · similarity 0.87)
- Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.87)
- Looking Beyond Single Images for Contrastive Semantic Segmentation Learning (NeurIPS 2021 · similarity 0.86)
- Perceptual Inductive Bias Is What You Need Before Contrastive Learning (CVPR 2025 · 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.
