Research map / Label-efficient and robust learning / Recognition and long-tail
Improved Deep Metric Learning with Multi-class N-pair Loss Objective
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 |
|---|---|---|---|
| replace triplet loss with multi-class N-pair loss (72 pairs per batch) | GoogLeNet embedding on Car-333 and Flower-610 · Car-333 / Flower-610 verification | VRF (neg=71): Car-333 48.96 to 76.02; Flower-610 73.04 to 88.63 | Helped |
| add negative class mining to triplet loss | triplet loss · Car-333 recognition | kNN recognition accuracy: 70.24 to 83.22 | Helped |
| use multi-class N-pair loss instead of one-vs-one N-pair loss (ablation) | N-pair loss · Car-333 / Flower-610 recognition and verification | kNN recognition and VRF (neg=71): Car-333 recognition 86.84 (ovo) to 88.37 (mc); VRF 73.05 to 76.02 | Helped |
| use multi-class N-pair loss instead of softmax loss | GoogLeNet embedding on Car-333 · Car-333 verification | VRF (neg=71): 55.36 to 76.02 | Helped |
| replace softmax with 72-pair-mc on Car-333 recognition | GoogLeNet embedding on Car-333 · Car-333 recognition | recognition accuracy: 88.69 (softmax) vs 88.37 (72-pair-mc), slightly worse | Hurt |
| multi-class N-pair loss with negative class mining on Online product | GoogLeNet embedding on Online product · Online product retrieval | recall@1: 53.32 (triplet) to 67.73 (60-pair-mc-nm) | Helped |
| add negative class mining to N-pair-mc loss (ablation) | 60-pair-mc on Online product · Online product retrieval | recall@1: 65.25 to 67.73 | Helped |
| multi-class N-pair loss without negative mining vs triplet loss | GoogLeNet embedding on Car-196 and CUB-200 · Car-196 / CUB-200 clustering and retrieval | recall@1: Car-196 53.84 to 71.12; CUB-200 43.30 to 50.96 | Helped |
| add negative class mining to triplet loss | triplet loss on Car-196 / CUB-200 · Car-196 / CUB-200 recall@1 | recall@1: Car-196 53.84 to 61.62; CUB-200 43.30 to 46.47 | Helped |
| increase number of classes per batch (N-pair) vs fewer classes with more examples per class (ablation) | 60x2 vs 10x12 N-pair batch construction on CUB-200 and Car-196 · Car-196 / CUB-200 / Online product recall@1 | recall@1: CUB-200 recall@1 49.65 (10x12) to 50.96 (60x2); Online product 63.58 (30x4) to 65.25 (60x2) | Helped |
| multi-class N-pair loss instead of triplet loss on LFW face verification | CasiaNet-style network trained on WebFace · LFW | verification accuracy: 95.88 to 98.27 (192-pair-mc); 98.33 with 320-pair-mc | Helped |
| multi-class N-pair loss instead of N-pair-ovo on LFW (ablation) | 192-pair loss on LFW · LFW | rank-1 identification: 66.21 (ovo) to 88.58 (mc) | Helped |
Closest papers by meaning
- Multi-Similarity Loss With General Pair Weighting for Deep Metric Learning (CVPR 2019 · similarity 0.89)
- OLÉ: Orthogonal Low-Rank Embedding - A Plug and Play Geometric Loss for Deep Learning (CVPR 2018 · similarity 0.88)
- Ranked List Loss for Deep Metric Learning (CVPR 2019 · similarity 0.88)
- Learning With Batch-Wise Optimal Transport Loss for 3D Shape Recognition (CVPR 2019 · similarity 0.87)
- Mean Field Theory in Deep Metric Learning (ICLR 2024 · similarity 0.86)
- Deep Disentangled Metric Learning (AAAI 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.