Research map / Image generation / Editing and style transfer
Towards Universal Image Embeddings: A Large-Scale Dataset and Challenge for Generic Image Representations
Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, Mário Lipovský, Pelin Dogan-Schönberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ondřej Chum, André Araujo
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 |
|---|---|---|---|
| train a single universal model on all 8 domains instead of an oracle assembly of specialists | CLIP ViT-B/16 64-D · UnED test (mean) | mMP@5: 59.6 -> 55.0 | Hurt |
| finetune on UnED training set instead of off-the-shelf 64-D PCA-whitened CLIP | CLIP ViT-B/16 · UnED test (mean) | mMP@5: 31.0 -> 55.0 | Improved |
| use PCA-whitening to reduce off-the-shelf embedding to 64-D (ablation) | CLIP and IN embeddings · UnED test (mean) | mMP@5: 39.8 -> 31.0 (CLIP) | Hurt |
| sample domains round-robin or by specialist steps instead of by dataset size (ablation) | universal CLIP training · UnED test (mean) | mMP@5: 52.4 -> 55.0 | Improved |
| use CLIP initialization instead of ImageNet for universal training (ablation) | universal embeddings · UnED test (mean) | mMP@5: 51.3 -> 55.0 (USCSS) | Improved |
| use a joint classifier instead of separate classifiers (ablation) | universal CLIP training with round-robin sampling · UnED test, SOP and InShop domains | R@1: InShop 89.4 (separate) -> 91.1 (joint) | Mixed |
| jointly train Met with GLDv2 in a universal model | universal embedding · UnED test, Met domain | mMP@5: 27.6 -> 2.5 | Hurt |
| finetune CLIP ViT-H/14 with stratified learning rates (projection/classifier 10^3 larger than backbone) | CLIP ViT-H/14 · Universal Embedding Challenge private split | mMP@5: 62.1 -> 70.9 | Improved |
| use bigger CLIP ViT-H/14 challenge solutions on UnED instead of paper baselines in some domains | challenge solutions · UnED test (CARS196, Met, GLDv2, Rp2k) | mMP@5/R@1 | Improved |
Closest papers by meaning
- MIEB: Massive Image Embedding Benchmark (ICCV 2025 · similarity 0.90)
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.89)
- UDON: Universal Dynamic Online distillatioN for generic image representations (NeurIPS 2024 · similarity 0.88)
- Illuminating Visual Identity in Universal Multimodal Embeddings (CVPR 2026 · similarity 0.87)
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.87)
- The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric (ICLR 2024 · 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.
