Research map / Image generation / Autoregressive and latent generation
Cross-Image-Attention for Conditional Embeddings in Deep Metric Learning
Read the paperCheck 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 cross-attention conditional embeddings, then use unconditional embeddings at inference | ResNet-50 DML with multi-similarity loss · CUB200-2011 test | Recall@1: 70.6 -> 74.6 | Helped |
| train with cross-attention conditional embeddings, then use unconditional embeddings at inference | ResNet-50 DML with multi-similarity loss · CARS196 test | Recall@1: 90.1 -> 91.2 | Helped |
| train with cross-attention conditional embeddings, then use unconditional embeddings at inference | ResNet-50 DML with multi-similarity loss · SOP test | Recall@1: 81.4 -> 82.3 | Helped |
| use conditional similarities s2 instead of unconditional s0 at inference (ablation) | Cross-Image-Attention · CUB200-2011 test | R@1: 74.6 -> 74.8 | No clear effect |
| use later conditional similarities s4, s6, or s8 instead of unconditional s0 at inference (ablation) | Cross-Image-Attention · CUB200-2011 test | R@1: 74.6 -> 74.1; 74.6 -> 73.9; 74.6 -> 74.0 | Hurt |
| use conditional similarities s2 instead of unconditional s0 at inference (ablation) | Cross-Image-Attention · CARS196 test | R@1: 91.2 -> 91.3 | No clear effect |
| use later conditional similarities s4, s6, or s8 instead of unconditional s0 at inference (ablation) | Cross-Image-Attention · CARS196 test | R@1: 91.2 -> 91.0; 91.2 -> 91.1; 91.2 -> 89.9 | No clear effect |
| increase the number of cross-attention blocks from 0 to multiple blocks (ablation) | Cross-Image-Attention · CUB200-2011 and CARS196 test | R@1: 66.4; 67.7; 70.2; 71.9; 72.2; 72.1; 74.6; 74.8; 74.1; 73.9; 74.0; 74.2; 74.3; 72.7; 72.8; 84.9; 86.7; 87.1; 89.0; 89.3; | Helped |
| replace the ResNet-50 encoder with ResNet-101 without cross-attention (ablation) | Multi-similarity DML · CUB200 and CARS196 test | R@1: 67.4% and 83.7% respectively; our model reaches 74.6% and 91.2% | Hurt |
Closest papers by meaning
- Divide and Conquer the Embedding Space for Metric Learning (CVPR 2019 · similarity 0.89)
- CrossFormer: A Versatile Vision Transformer Hinging on Cross-scale Attention (ICLR 2022 · similarity 0.88)
- Integrating Language Guidance Into Vision-Based Deep Metric Learning (CVPR 2022 · similarity 0.88)
- Cross-Batch Memory for Embedding Learning (CVPR 2020 · similarity 0.87)
- Potential Field Based Deep Metric Learning (CVPR 2025 · similarity 0.86)
- Learning Semantic Proxies from Visual Prompts for Parameter-Efficient Fine-Tuning in Deep Metric Learning (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.