Research map / Label-efficient and robust learning / Recognition and long-tail
Contrastive Multiview Coding
Yonglong Tian, Dilip Krishnan, Phillip Isola
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 |
|---|---|---|---|
| use Y,DbDr colorspace instead of L,ab | CMC ResNet-50x2 · ImageNet-1k | top-1 accuracy: 68.3 / 88.2 -> 69.0 / 88.9 | Improved |
| add RandAugment (ablation) | CMC {Y,DbDr} ResNet-50x2 · ImageNet-1k | top-1 accuracy: 69.0 -> 70.6 | Improved |
| add dorsal (optical flow) stream to ventral stream (ablation) | CMC CaffeNet · UCF-101 | accuracy: 55.3 -> 59.1 | Improved |
| add dorsal (optical flow) stream to ventral stream | CMC CaffeNet · HMDB-51 | accuracy: 26.7 vs 27.2 for Cross and Learn | No clear effect |
| add more views (depth, normals) beyond L and ab (ablation) | CMC patch-based · NYU-Depth-V2 segmentation | mIoU | Improved |
| full-graph instead of core-view paradigm (ablation) | CMC · NYU-Depth-V2 segmentation from L | mIoU: 34.1 -> 34.4 | No clear effect |
| use contrastive instead of predictive learning (ablation) | L,Depth encoder · STL-10 | accuracy: 55.5 -> 58.3 | Improved |
| use contrastive instead of predictive learning (ablation) | L,Normal encoder · STL-10 | accuracy: 58.4 -> 60.1 | Improved |
| use contrastive instead of predictive learning (ablation) | L,Seg. Map encoder · STL-10 | accuracy: 57.7 -> 59.2 | Improved |
| increase number of negatives from 64 to 4096 (ablation) | CMC AlexNet · ImageNet 100 classes | accuracy: saturates at around 60.3% with m=4096 | Improved |
| add MoCo momentum encoder (ablation) | CMC · ImageNet-100 | top-1 accuracy: 75.6 -> 77.2 (+1.6) | Improved |
| add PIRL JigSaw branch (ablation) | CMC · ImageNet-100 | top-1 accuracy: 75.6 -> 78.0 (+2.4) | Improved |
Closest papers by meaning
- What Should Not Be Contrastive in Contrastive Learning (ICLR 2021 · similarity 0.88)
- Poly-View Contrastive Learning (ICLR 2024 · similarity 0.87)
- Learning from Unlabelled Videos Using Contrastive Predictive Neural 3D Mapping (ICLR 2020 · similarity 0.87)
- Perceptual Inductive Bias Is What You Need Before Contrastive Learning (CVPR 2025 · similarity 0.86)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.86)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 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.
