Research map / Vision-language and multimodal / CLIP and zero-shot
Unified Contrastive Learning in Image-Text-Label Space
Check 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 |
|---|---|---|---|
| Train with UniCL instead of CrossEntropy | ResNet-50 image classification · CIFAR-10 | Top-1 accuracy: 95.0 -> 96.8 | Helped |
| Train with UniCL instead of CrossEntropy | Swin-Tiny image classification · ImageNet-1K | Top-1 accuracy: 76.8 -> 79.9 | Helped |
| Train with UniCL instead of SupCon | ResNet-50 image classification · ImageNet-1K | Top-1 accuracy: 78.7 -> 78.1 | Hurt |
| Add image-text pairs to ImageNet-1K pretraining using UniCL | Swin-Tiny UniCL · Zero-shot 14 datasets | Score: 30.2 -> 45.1 | Helped |
| Add GCC-3M to ImageNet-1K pretraining | Swin-Tiny UniCL · COCO detection | Box mAP: 42.5 -> 43.0 | Helped |
| Add YFCC-14M to ImageNet-1K pretraining | Swin-Tiny UniCL · 18 datasets | Linear-probe score: 78.0 -> 80.1 | Helped |
| Use GCC-15M rather than GCC-3M with ImageNet-1K | Swin-Tiny UniCL · Zero-shot 14 datasets | Score: 39.0 -> 45.1 | Helped |
| Combine YFCC-14M and ImageNet-21K in UniCL rather than train CLIP on YFCC-14M | Swin-Tiny · Zero-shot 14 datasets | Score: 36.3 -> 45.5 | Helped |
| Add the bidirectional image-to-text loss (ablation) | UniCL with embedding text encoder · ImageNet-1K | Top-1 accuracy: 75.7 -> 78.7 | Helped |
| Use a Transformer text encoder instead of a simple embedding layer (ablation) | UniCL · ImageNet-1K | Top-1 accuracy: 78.7 -> 79.9 | Helped |
| Reduce batch size from 4096 to 2048 or 1024 (ablation) | UniCL · ImageNet-1K | Top-1 accuracy: Transformer: 79.9 -> 80.1; 79.9; Embedding: 78.7 -> 78.9; 79.0 | No clear effect |
| Add image-label data to image-text pretraining using UniCL | Swin-Tiny model pretrained on YFCC-14M + ImageNet-21K (half/half) · ImageNet-1K zero-shot | Accuracy: +6% | Helped |
Closest papers by meaning
- Language-Driven Cross-Modal Classifier for Zero-Shot Multi-Label Image Recognition (ICML 2024 · similarity 0.87)
- Multi-Label Supervised Contrastive Learning (AAAI 2024 · similarity 0.86)
- FFF: Fixing Flawed Foundations in Contrastive Pre-Training Results in Very Strong Vision-Language Models (CVPR 2024 · similarity 0.86)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.86)
- Beyond Words: Augmenting Discriminative Richness via Diffusions in Unsupervised Prompt Learning (CVPR 2025 · similarity 0.86)
- ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 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.