Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Deep Descriptors With Scale-Aware Triplet Networks
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 |
|---|---|---|---|
| add mixed global and triplet-local context (γ=0.5) to scale-aware triplet training (ablation) | L2-Net · UBC benchmark | FPR95: 3.04 -> 2.96; 0.64 -> 0.68; 1.61 -> 1.79; 1.02 -> 1.02; 3.15 -> 2.51; 2.21 -> 1.64 | Mixed |
| add scale-aware sampling (ablation) | Triplet L2-Net on UBC · UBC benchmark | FPR95: Random sampling, δ = 12.5 versus scale-aware sampling, δ = 5; e.g. 9.72 -> 3.04 | Helped |
| add scale correction (δ=12.5) to the log loss (ablation) | Triplet L2-Net with random sampling · UBC benchmark | FPR95: Table 1: δ=1 versus δ=12.5; e.g. 45.0 -> 4.82 | Helped |
| remove scale correction (use δ=1) (ablation) | Triplet L2-Net with random sampling · UBC benchmark | FPR95: 45.0, 43.2, 37.2, 37.1, 66.0, 62.6 versus 4.82, 9.72, 1.90, 3.55, 10.5, 6.87 | Hurt |
| replace the triplet loss with a Siamese loss (γ=0) (ablation) | Scale-aware L2-Net · UBC benchmark | FPR95: 4.70, 1.36, 2.61, 1.62, 3.89, 2.91 versus triplet 3.04, 0.64, 1.61, 1.02, 3.15, 2.21 | Mixed |
| add mixed-context loss (γ=0.5) instead of pure Siamese loss (ablation) | Scale-aware L2-Net · MNIST test set | FPR95: 7.6% -> 4.8% | Helped |
| replace mixed-context loss with pure triplet loss (γ=1) (ablation) | Scale-aware L2-Net · MNIST test set | FPR95: 4.8% -> 12.5% | Hurt |
| vary scale correction parameter δ (ablation) | Triplet L2-Net with random sampling · UBC liberty→notredame | FPR95 | Mixed |
| add scale-aware sampling rather than random sampling (ablation) | Triplet L2-Net · UBC liberty→notredame | FPR95 | Helped |
| use mixed-context descriptors | L2-Net · HPatches verification | mAP: 85.75% -> 86.35% | Helped |
| use mixed-context descriptors | L2-Net · HPatches matching | mAP: 46.32% -> 48.22% | Helped |
| use mixed-context descriptors | L2-Net · HPatches retrieval | mAP: 55.37% -> 57.54% | Helped |
Closest papers by meaning
- Learning Local Image Descriptors With Deep Siamese and Triplet Convolutional Networks by Minimising Global Loss Functions (CVPR 2016 · similarity 0.89)
- Working hard to know your neighbor's margins: Local descriptor learning loss (NeurIPS 2017 · similarity 0.87)
- Triplet-Center Loss for Multi-View 3D Object Retrieval (CVPR 2018 · similarity 0.86)
- HyNet: Learning Local Descriptor with Hybrid Similarity Measure and Triplet Loss (NeurIPS 2020 · similarity 0.86)
- Learning With Batch-Wise Optimal Transport Loss for 3D Shape Recognition (CVPR 2019 · similarity 0.85)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.84)
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.