Research map / Label-efficient and robust learning / Recognition and long-tail
Better and Faster: Exponential Loss for Image Patch Matching
Shuang Wang, Yanfeng Li, Xuefeng Liang, Dou Quan, Bowu Yang, Shaowei Wei, Licheng Jiao
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 linear triplet loss with exponential triplet loss (metric learning) (ablation) | HardNet-style network · UBC | mean FPR95: 1.56 -> 0.98 | Improved |
| replace linear triplet loss with exponential triplet loss (descriptor learning) (ablation) | HardNet · UBC | mean FPR95: 1.25 -> 1.00 | Improved |
| replace linear with exponential losses in HardNet (ablation) | HardNet · UBC | training time | Improved |
| add exponential Siamese loss instead of linear on MNIST toy (ablation) | Siamese network · MNIST clustering | STD of intra-class distances: 0.71 -> 0.57 | Improved |
| add exponential triplet loss instead of linear on MNIST toy (ablation) | triplet network · MNIST clustering | STD of intra-class distances: 0.81 -> 0.65 | Improved |
| add hard positive mining at ratio 1:2 (batch size 128) (ablation) | Exp-TLoss (dist) · UBC | FPR95: 1.09 -> 1.03 | Improved |
| use exponential orders beta=2, gamma=0.3 rather than linear orders (ablation) | Exp-TLoss (prob) · UBC (Liberty-trained) | FPR95: 0.75 | Improved |
| train on full UBC dataset (FB) instead of Liberty only | Exp-TLoss (dist) · HPatches patch verification | mAP: 87.70% -> 90.07% | Improved |
| use Exp-TLoss (dist) instead of HardNet+ when trained on Liberty | descriptor learning · HPatches patch verification | mAP: 88.43% -> 87.70% | Hurt |
| use Exp-TLoss (dist) instead of HardNet+ on RGB-NIR | descriptor learning · RGB-NIR | mean FPR95: 2.80 -> 2.07 | Improved |
| use Exp-TLoss (prob) instead of SNet on Notredame-trained UBC | metric learning · UBC | FPR95 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- L2-Net — L2-Net: Deep Learning of Discriminative Patch Descriptor in Euclidean Space (CVPR 2017) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AFD-Net: Aggregated Feature Difference Learning for Cross-Spectral Image Patch Matching (ICCV 2019 · similarity 0.85)
- On the Convergence of PatchMatch and Its Variants (CVPR 2018 · similarity 0.85)
- Hard Patches Mining for Masked Image Modeling (CVPR 2023 · similarity 0.85)
- Hard negative examples are hard, but useful (ECCV 2020 · similarity 0.85)
- CNN-Based Patch Matching for Optical Flow With Thresholded Hinge Embedding Loss (CVPR 2017 · similarity 0.84)
- Mathematical Justification of Hard Negative Mining via Isometric Approximation Theorem (ICLR 2024 · similarity 0.83)
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.
