Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Spread-Out Local Feature Descriptors
Xu Zhang, Felix X. Yu, Sanjiv Kumar, Shih-Fu Chang
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 |
|---|---|---|---|
| add global orthogonal regularization (GOR) to contractive loss, versus best pairwise model DeepCompare 2str | contractive loss · UBC patch mean FPR95 | FPR95 (%): 9.67 -> 5.97 | Improved |
| add GOR to triplet loss | TFeat triplet loss · UBC patch mean FPR95 | FPR95 (%): 6.79 -> 4.69 | Improved |
| add GOR to triplet loss with anchor swap | TFeat+AS · UBC patch mean FPR95 | FPR95 (%): 6.47 -> 4.36 | Improved |
| add GOR to N-pair loss | N-pair · UBC patch mean FPR95 | FPR95 (%): 5.29 -> 4.98 | Improved |
| add anchor swap on top of triplet loss with GOR | TL+GOR · UBC patch mean FPR95 | FPR95 (%): 4.69 -> 4.36 | Improved |
| use hinge on the second-moment term of GOR, versus l1 or l2 (ablation) | GOR · UBC patch | FPR95 (%) | No clear effect |
| use small descriptor dimension d=32 with GOR (ablation) | GOR with triplet loss · UBC patch, embedding dimension | FPR95 (%) | Hurt |
| use GOR with descriptor dimension d >= 64 (ablation) | GOR with triplet loss · UBC patch, embedding dimension | FPR95 (%) | Improved |
| add GOR to LSSS loss for image embedding | LSSS · Stanford Online Products | F1, NMI, Recall@K | Improved |
| add GOR to metric learning on Car196 and CUB-200-2011 | embedding loss · Car196 and CUB-200-2011 | Recall | No clear effect |
| add GOR, measured by overlap of similarity histograms | triplet loss · UBC Notre Dame -> Liberty | empirical error (histogram overlap): 15% relative decrease | Improved |
Closest papers by meaning
- Local Descriptors Optimized for Average Precision (CVPR 2018 · similarity 0.86)
- Working hard to know your neighbor's margins: Local descriptor learning loss (NeurIPS 2017 · similarity 0.86)
- Learning Local Image Descriptors With Deep Siamese and Triplet Convolutional Networks by Minimising Global Loss Functions (CVPR 2016 · similarity 0.85)
- Learning Deep Descriptors With Scale-Aware Triplet Networks (CVPR 2018 · similarity 0.85)
- Learning Local Descriptors With a CDF-Based Dynamic Soft Margin (ICCV 2019 · similarity 0.85)
- Boosting Few-Shot Learning via Attentive Feature Regularization (AAAI 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.
