Research map / Architectures and efficiency / CNN design
Working hard to know your neighbor's margins: Local descriptor learning loss
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 L2Net's learning objective with hardest-in-batch sampling and triplet margin loss | L2Net descriptor · Brown mean FPR95 | FPR95: 3.24 -> 1.57 | Helped |
| Use HardNet rather than L2Net+ | Local descriptor · Oxford5k BoW+QE | mAP: 80.4 -> 83.2 | Helped |
| Use HardNet rather than L2Net+ | Local descriptor · Paris6k BoW+SV | mAP: 66.6 -> 67.4 | Helped |
| Train on Brown and HPatches instead of Liberty only | HardNet · Oxford5k BoW+QE | mAP: 83.0 -> 84.5 | Helped |
| Train on Brown and HPatches instead of Liberty only | HardNet · Paris6k BoW+QE | mAP: 77.5 -> 79.1 | Helped |
| Add hardest-in-batch sampling rather than random sampling (ablation) | HardNet training · HPatches matching | mean mAP: 0.349 -> 0.482 | Helped |
| Add hardest-in-batch sampling rather than epoch-level hard negative mining (ablation) | HardNet training · HPatches matching | mean mAP: 0.346 -> 0.482 | Helped |
| Use hardest-in-batch sampling without CPR rather than random sampling with CPR (ablation) | HardNet training · HPatches matching | mean mAP: 0.286 -> 0.482 | Helped |
| Use triplet margin loss rather than softmin with hardest-in-batch sampling (ablation) | HardNet training · HPatches matching | mean mAP: 0.474 -> 0.482 | Helped |
| Use triplet margin loss rather than contrastive margin 1 with hardest-in-batch sampling (ablation) | HardNet training · HPatches matching | mean mAP: 0.444 -> 0.482 | Helped |
| Increase batch size up to 512 (ablation) | HardNet training · Brown validation sequences | FPR95 | Helped |
| Increase batch size beyond 512 (ablation) | HardNet training · Brown validation sequences | FPR95 | No clear effect |
Closest papers by meaning
- Learning Deep Descriptors With Scale-Aware Triplet Networks (CVPR 2018 · similarity 0.87)
- Understanding and Simplifying Perceptual Distances (CVPR 2021 · similarity 0.86)
- Learning With Batch-Wise Optimal Transport Loss for 3D Shape Recognition (CVPR 2019 · similarity 0.86)
- Local Descriptors Optimized for Average Precision (CVPR 2018 · similarity 0.86)
- Improved Deep Metric Learning with Multi-class N-pair Loss Objective (NeurIPS 2016 · similarity 0.86)
- Learning Local Image Descriptors With Deep Siamese and Triplet Convolutional Networks by Minimising Global Loss Functions (CVPR 2016 · 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.