Research map / Vision-language and multimodal / CLIP and zero-shot
Retrieval Augmented Classification for Long-Tail Visual Recognition
Alexander Long, Wei Yin, Thalaiyasingam Ajanthan, Vu Nguyen, Pulak Purkait, Ravi Garg, Alan Blair, Chunhua Shen, Anton van den Hengel
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 |
|---|---|---|---|
| add the retrieval branch to the LACE-trained base encoder (ablation) | ViT-B-16 on Places365-LT · Places365-LT | top-1 accuracy, All: 44.05 -> 47.17 | Improved |
| add the retrieval branch to the LACE-trained base encoder (ablation) | ViT-B-16 on iNaturalist 2018 · iNaturalist 2018 validation | top-1 accuracy, All: 78.32 -> 80.24 | Improved |
| add the retrieval branch to the LACE-trained base encoder (ablation) | ViT-B-16 on Places365-LT · Places365-LT | top-1 accuracy, Few: 40.77 -> 41.76 | Improved |
| add the retrieval branch to the LACE-trained base encoder (ablation) | ViT-B-16 on iNaturalist 2018 · iNaturalist 2018 validation | top-1 accuracy, Few: 78.55 -> 81.07 | Improved |
| use a ViT-B-16 rather than a ViT-B-32 retrieval encoder (ablation) | retrieval branch · iNaturalist 2018 retrieval | top-1 accuracy, All: 44.31 -> 53.89 | Improved |
| increase retrieval input resolution from 224×224 to 384×384 (ablation) | ViT-B-16 retrieval branch · iNaturalist 2018 retrieval | top-1 accuracy, All: 53.89 -> 61.77 | Improved |
| replace exact k-NN with HNSW approximate k-NN (ablation) | Places365-LT retrieval index, 62.5k items, L2 distance · Places365-LT retrieval | query time: 1.12 -> 1.06 | Improved |
| replace exact k-NN with HNSW approximate k-NN (ablation) | Places365-LT retrieval index, 62.5k items, L2 distance · Places365-LT retrieval | top-1 accuracy: 29.91 -> 29.66 | Hurt |
| replace L2 distance with cosine distance (ablation) | exact k-NN retrieval · Places365-LT retrieval, 62.5k index | top-1 accuracy: 29.91 -> 29.64 | Hurt |
| increase retrieval input resolution from 224×224 to 384×384 (ablation) | ViT-B-16 retrieval branch · Places365-LT retrieval | top-1 accuracy, All: 29.91 -> 30.55 | Improved |
| add inverse-square-root class-frequency reweighting to logit adjustment (ablation) | RAC · Places365-LT | top-1 accuracy, All: 47.75 -> 48.32 | Improved |
| add inverse-log class-frequency reweighting to logit adjustment (ablation) | RAC · Places365-LT | top-1 accuracy, All: 47.75 -> 48.29 | Improved |
Closest papers by meaning
- Improving Image Recognition by Retrieving From Web-Scale Image-Text Data (CVPR 2023 · similarity 0.91)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.87)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.86)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.86)
- Learning from Memory: Non-Parametric Memory Augmented Self-Supervised Learning of Visual Features (ICML 2024 · similarity 0.86)
- Retrieval-Augmented Open-Vocabulary Object Detection (CVPR 2024 · 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.
