Research map / Detection and segmentation / Open-set and few-shot detection
Soft Neighbors are Positive Supporters in Contrastive Visual Representation Learning
Chongjian GE, Jiangliu Wang, Zhan Tong, Shoufa Chen, Yibing Song, Ping Luo
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 soft positiveness weights to neighbors (vs. treating neighbors as exactly positive) (ablation) | SNCLR with 30 neighbors, ViT-S · ImageNet linear eval (300 ep) | top-1 accuracy: +1.14 (72.24 to 73.38) | Improved |
| add nearest-neighbor soft positives (K=30) vs. no neighbors (ablation) | SNCLR, ViT-S · ImageNet linear eval (300 ep) | top-1 accuracy: +1.07 (71.17 to 72.24) without positiveness; +2.21 with positiveness | Improved |
| neighbor number K from 30 to 10 or 50 (ablation) | SNCLR, ViT-S · ImageNet linear eval (300 ep) | top-1 accuracy: 72.81 (K=10) and 72.92 (K=50) vs 73.38 (K=30) | Hurt |
| increase candidate neighbor set length from 8000 to 128000 (ablation) | SNCLR, ViT-S · ImageNet linear eval (300 ep) | top-1 accuracy: 72.25 to 73.38 (non-monotonic, e.g. 32000 gives 72.43) | Improved |
| apply soft neighbors to positives only vs. negatives only vs. both (ablation) | SNCLR, ViT-S · ImageNet linear eval (300 ep) | top-1 accuracy: positives only +0.71, negatives only +1.26, both +2.21 over baseline 71.17 | Improved |
| SNCLR (ResNet-50, 800 ep) vs. prior SSL methods | ResNet-50 · ImageNet linear eval | top-1 accuracy: 75.3 vs 74.5 (NNCLR) and 74.3 (BYOL) | Improved |
| SNCLR vs. baselines on ViT-S and ViT-B | ViT-S, ViT-B · ImageNet linear eval (300 ep) | top-1 accuracy: 73.4 vs 72.5 (DINO) on ViT-S; 76.8 vs 76.2 (MoCo v3) on ViT-B | Improved |
| SNCLR vs. baselines on COCO detection and instance segmentation | ResNet-50 Mask R-CNN · COCO det./seg. 1x | AP^bb / AP^mk: 39.4 AP^bb at 200 ep equals BYOL-800; 39.9 at 800 ep vs BYOL 39.4 | Improved |
| cross-attention module uses learned linear projections instead of identity mappings | positiveness attention module · ImageNet linear eval (ViT-S) | top-1 accuracy | No clear effect |
Closest papers by meaning
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.88)
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.87)
- Demystifying Contrastive Self-Supervised Learning: Invariances, Augmentations and Dataset Biases (NeurIPS 2020 · similarity 0.87)
- Revisiting Contrastive Learning through the Lens of Neighborhood Component Analysis: an Integrated Framework (ICML 2022 · similarity 0.87)
- Mutual Contrastive Learning for Visual Representation Learning (AAAI 2022 · similarity 0.87)
- Single-View Graph Contrastive Learning with Soft Neighborhood Awareness (AAAI 2025 · similarity 0.86)
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.
