Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Local Descriptors With a CDF-Based Dynamic Soft Margin
Linguang Zhang, Szymon Rusinkiewicz
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 hard-margin triplet loss with CDF-based dynamic soft margin | HardNet+ · UBC PhotoTourism mean, real-valued | FPR95 (%): 1.52 -> 1.18 | Improved |
| replace hard-margin triplet loss with CDF-based dynamic soft margin | L2-Net binary · UBC PhotoTourism mean, binary | FPR95 (%): 3.01 -> 2.63 | Improved |
| use dynamic soft margin instead of softplus static soft margin (ablation) | triplet loss · Real-valued, UBC FPR95 | FPR95 (%): 1.20 -> 0.95 | Improved |
| use dynamic soft margin instead of Wang et al. dynamic hard margin (ablation) | triplet loss · Real-valued, UBC FPR95 | FPR95 (%): 1.18 -> 0.95 | Improved |
| use dynamic soft margin instead of hardest 1/8 mining (ablation) | triplet loss · Real-valued, UBC FPR95 | FPR95 (%): 2.19 -> 0.95 | Improved |
| use dynamic soft margin instead of softplus (ablation) | triplet loss · Binary, HPatches patch retrieval | mAP (%): 54.23 -> 55.12 | Improved |
| build the PDF from dpos instead of dpos - dneg (ablation) | dynamic soft margin · Real-valued, UBC FPR95 | FPR95 (%): 0.95 -> 0.98 | No clear effect |
| build the PDF from dneg instead of dpos - dneg (ablation) | dynamic soft margin · Binary, UBC FPR95 | FPR95 (%): 2.31 -> 3.02 | Hurt |
| approximate PDF with a Gaussian instead of histogram (ablation) | dynamic soft margin · Real-valued, UBC FPR95 | FPR95 (%): 0.95 -> 1.07 | Hurt |
| train on union of UBC PhotoTourism and HPatches (Ours++) | dynamic soft margin · Oxford5k HQE multiple assignment | mAP (%): 85.94 -> 86.42 | Improved |
| tune triplet margin from the default 1.0 to 0.375 (ablation) | HardNet · UBC real-valued descriptor | FPR95 | Improved |
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 2 of 2 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
- HyNet: Learning Local Descriptor with Hybrid Similarity Measure and Triplet Loss (NeurIPS 2020 · similarity 0.86)
- Descriptor Distillation: A Teacher-Student-Regularized Framework for Learning Local Descriptors (IJCV 2024 · similarity 0.86)
- Learning Deep Descriptors With Scale-Aware Triplet Networks (CVPR 2018 · similarity 0.86)
- Learning Local Image Descriptors With Deep Siamese and Triplet Convolutional Networks by Minimising Global Loss Functions (CVPR 2016 · similarity 0.86)
- Revisiting Unsupervised Local Descriptor Learning (AAAI 2023 · similarity 0.85)
- SDGMNet: Statistic-Based Dynamic Gradient Modulation for Local Descriptor Learning (AAAI 2024 · 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.
