Research map / Label-efficient and robust learning / Recognition and long-tail
Un-mix: Rethinking Image Mixtures for Unsupervised Visual Representation 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 Un-Mix mixtures and softened pair-distance losses | MoCo V2 · ImageNet, 200 epochs | Top-1 accuracy (%): 67.5 -> 68.6; 1.1% | Helped |
| Add Un-Mix mixtures and softened pair-distance losses | MoCo V2 · ImageNet, 800 epochs | Top-1 accuracy (%): 71.1 -> 71.8; ↑0.7 | Helped |
| Add Un-Mix mixtures and softened pair-distance losses | BYOL · CIFAR-10 | linear accuracy (%): 91.73 -> 94.20 | Helped |
| Add Un-Mix mixtures and softened pair-distance losses | BYOL · CIFAR-100 | linear accuracy (%): 66.60 -> 71.50 | Helped |
| Add Un-Mix mixtures and softened pair-distance losses | BYOL · STL-10 | linear accuracy (%): 91.99 -> 93.34 | Helped |
| Add Un-Mix mixtures and softened pair-distance losses | BYOL · Tiny ImageNet | linear accuracy (%): 51.00 -> 53.39 | Helped |
| Pretrain MoCo V2 with Un-Mix | MoCo V2 · PASCAL VOC test2007, Faster R-CNN, 200-epoch pretraining | AP50: 82.4 -> 83.0; ↑0.6 | Helped |
| Pretrain MoCo V2 with Un-Mix | MoCo V2 · COCO val2017, Mask R-CNN, 200-epoch pretraining | AP: 40.9 -> 41.2; ↑0.3 | Helped |
| Add the original-image loss alongside mixed-pair losses (ablation) | MoCo V2 + mixed-pair losses · ImageNet, 200 epochs | Top-1 accuracy (%): 66.8 -> 68.6 | Helped |
| Remove mixed-pair loss terms (ablation) | MoCo V2 + Un-Mix · ImageNet, 200 epochs | Top-1 accuracy (%): 68.6 -> 67.5 | Hurt |
| Use region-level mixtures only instead of global mixtures only (ablation) | MoCo V2 + Un-Mix · ImageNet | Top-1 accuracy (%): 67.6 -> 68.6 | Helped |
| Set global-mixture probability to 0.5 rather than 1.0 (ablation) | MoCo V2 + Un-Mix · ImageNet | Top-1 accuracy (%): 67.6 -> 68.3 | Helped |
Closest papers by meaning
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.90)
- Deep Unsupervised Similarity Learning Using Partially Ordered Sets (CVPR 2017 · similarity 0.87)
- Active Learning by Feature Mixing (CVPR 2022 · similarity 0.87)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.86)
- Beyond Words: Augmenting Discriminative Richness via Diffusions in Unsupervised Prompt Learning (CVPR 2025 · similarity 0.86)
- Integrating Low-Level Visual Cues for Enhanced Unsupervised Semantic Segmentation (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.