Research map / Vision-language and multimodal / Human and embodied modeling
Improving Image Recognition by Retrieving From Web-Scale Image-Text Data
Ahmet Iscen, Alireza Fathi, Cordelia Schmid
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 |
|---|---|---|---|
| Learn attention weights for retrieved examples instead of averaging their values | Retrieval-augmented classifier · ImageNet-LT | accuracy (All): 75.8 -> 78.3 | Improved |
| Learn attention weights for retrieved examples instead of averaging their values | Retrieval-augmented classifier · Places-LT | accuracy (All): 44.9 -> 49.9 | Improved |
| Learn attention weights for retrieved examples instead of averaging their values | Retrieval-augmented classifier · iNaturalist2021-Mini | accuracy: 58.9 -> 66.2 | Improved |
| Learn attention weights for retrieved examples instead of averaging their values | Retrieval-augmented classifier · Webvision | accuracy: 78.2 -> 83.6 | Improved |
| Fine-tune the visual encoder along with the memory attention module (ablation) | MAM · ImageNet-LT | accuracy (All): 78.3 -> 82.3 | Improved |
| Fine-tune the visual encoder along with the memory attention module (ablation) | MAM · Places-LT | accuracy (All): 49.9 -> 51.4 | Improved |
| Use T5-Base text embeddings rather than visual memory-value embeddings (ablation) | MAM · ImageNet-LT | accuracy | Improved |
| Increase the memory dataset size (ablation) | MAM · ImageNet-LT | accuracy | Improved |
| Increase the number of retrieved neighbors on a small memory such as ImageNet-LT (ablation) | MAM with T5-Base memory values · ImageNet-LT | accuracy | Hurt |
| Increase the number of retrieved neighbors on a large memory dataset (ablation) | MAM with T5-Base memory values · YFCC, LAION, and WebLI | accuracy | No clear effect |
| Add more image-text pairs to the memory after training without retraining MAM (ablation) | MAM · ImageNet-LT | accuracy | Improved |
Closest papers by meaning
- Retrieval Augmented Classification for Long-Tail Visual Recognition (CVPR 2022 · similarity 0.91)
- Thinking Fast and Slow: Efficient Text-to-Visual Retrieval With Transformers (CVPR 2021 · similarity 0.87)
- Learning from Memory: Non-Parametric Memory Augmented Self-Supervised Learning of Visual Features (ICML 2024 · similarity 0.87)
- Retrieval-Enhanced Contrastive Vision-Text Models (ICLR 2024 · similarity 0.86)
- Highlighting What Matters: Promptable Embeddings for Attribute-Focused Image Retrieval (NeurIPS 2025 · similarity 0.86)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (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.
