Research map / Label-efficient and robust learning / Recognition and long-tail
Sampling Matters in Deep Embedding Learning
Chao-Yuan Wu, R. Manmatha, Alexander J. Smola, Philipp Krahenbuhl
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 |
|---|---|---|---|
| replace random sampling with semi-hard negative sampling (ablation) | contrastive loss · Stanford Online Products (from scratch, ablation) | Recall@1: 30.1 -> 49.4 | Improved |
| replace random sampling with distance weighted sampling (ablation) | margin loss · Stanford Online Products (from scratch, ablation) | Recall@1: 37.5 -> 61.7 | Improved |
| replace semi-hard sampling with distance weighted sampling (ablation) | margin loss · Stanford Online Products (from scratch, ablation) | Recall@1: 61.0 -> 61.7 | Improved |
| replace semi-hard sampling with distance weighted sampling (ablation) | triplet l2 · Stanford Online Products (from scratch, ablation) | Recall@1: 47.4 -> 54.5 | Improved |
| replace semi-hard sampling with distance weighted sampling (ablation) | contrastive loss · Stanford Online Products (from scratch, ablation) | Recall@1: 49.4 -> 39.2 | Hurt |
| replace triplet l2 loss with margin loss (ablation) | distance weighted sampling · Stanford Online Products (from scratch, ablation) | Recall@1: 54.5 -> 61.7 | Improved |
| replace contrastive loss with margin loss (ablation) | semi-hard sampling · Stanford Online Products (from scratch, ablation) | Recall@1: 49.4 -> 61.0 | Improved |
| pre-train on ILSVRC 2012-CLS (ablation) | margin loss with distance weighted sampling · Stanford Online Products (from scratch, ablation) | Recall@1: 61.7 -> 72.7 | Improved |
| learn class-specific beta instead of fixed beta=1.2 (ablation) | margin loss · Stanford Online Products (pre-trained, 8K iterations) | Recall@1: 71.1 -> 72.7 | Improved |
| use fixed beta=0.8 instead of 1.2 (ablation) | margin loss · Stanford Online Products (pre-trained, 8K iterations) | Recall@1: 71.1 -> 61.3 | Hurt |
| use example-specific beta (ablation) | margin loss · Stanford Online Products (pre-trained, 8K iterations) | Recall@1 | No clear effect |
| use margin loss with distance weighted sampling instead of HDC | image retrieval · CARS196 (original images) | Recall@1: 73.7 -> 79.6 | Improved |
Closest papers by meaning
- Deep Embedding Learning With Discriminative Sampling Policy (CVPR 2019 · similarity 0.89)
- DAS: Densely-Anchored Sampling for Deep Metric Learning (ECCV 2022 · similarity 0.88)
- Sampling Wisely: Deep Image Embedding by Top-K Precision Optimization (ICCV 2019 · similarity 0.88)
- $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.87)
- Local Similarity-Aware Deep Feature Embedding (NeurIPS 2016 · similarity 0.87)
- Stabilizing Sample Similarity in Representation via Mitigating Random Consistency (ICML 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.
