Research map / Label-efficient and robust learning / Recognition and long-tail
Unicom: Universal and Compact Representation Learning for Image Retrieval
Xiang An, Jiankang Deng, Kaicheng Yang, Jaiwei Li, Ziyong Feng, Jia Guo, Jing Yang, Tongliang Liu
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 cluster discrimination on LAION 400M instead of instance discrimination pre-training | CLIP/OPEN-CLIP visual encoders · Unsupervised retrieval (average across 7 datasets) | average retrieval score: 64.5 -> 69.9 | Improved |
| Use joint image-text features rather than image features alone for clustering (ablation) | UNICOM · Linear probe (average, 13 datasets) | top-1 accuracy: 84.4 -> 85.7 | Improved |
| Use joint image-text features rather than image features alone for clustering (ablation) | UNICOM · Unsupervised retrieval (average) | retrieval score: 61.8 -> 62.8 | Improved |
| Use 1M clusters rather than 100K clusters (ablation) | UNICOM · Linear probe (average, 13 datasets) | top-1 accuracy: 75.9 -> 85.7 | Improved |
| Use 1M clusters rather than 10M clusters (ablation) | UNICOM · Linear probe (average, 13 datasets) | top-1 accuracy: 83.6 -> 85.7 | Improved |
| Sample only a subset of negative classes (r1=0.1) rather than use all classes (ablation) | UNICOM · Linear probe (average, 13 datasets) | top-1 accuracy: 77.9 -> 85.7 | Improved |
| Lower the random class ratio from 0.1 to 0.05 (ablation) | UNICOM · Linear probe (average, 13 datasets) | top-1 accuracy: 85.7 -> 85.1 | Hurt |
| Apply shared random feature selection with r2=0.5 instead of full dimensions (ablation) | UNICOM · Unsupervised retrieval at 256 dimensions | retrieval score: 61.4 -> 61.8 | Improved |
| Replace shared random feature selection with Dropout (ablation) | UNICOM · Unsupervised retrieval at 256 dimensions | retrieval score: 61.8 -> 60.7 | Hurt |
| Train from scratch rather than fine-tune from CLIP (ablation) | UNICOM ViT B/32 · 13-dataset linear probe (average) | top-1 accuracy: 85.0 -> 85.7 | Improved |
| Add random feature selection at full 512-dimensional representation (ablation) | UNICOM · Linear probe and full-dimension unsupervised retrieval | performance | No clear effect |
Closest papers by meaning
- Coarse-to-Fine: Learning Compact Discriminative Representation for Single-Stage Image Retrieval (ICCV 2023 · similarity 0.89)
- U-MARVEL: Unveiling Key Factors for Universal Multimodal Retrieval via Embedding Learning with MLLMs (ICLR 2026 · similarity 0.86)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.86)
- NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.86)
- Unsupervised Cross-Domain Image Retrieval via Prototypical Optimal Transport (AAAI 2024 · similarity 0.86)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 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.
