Research map / Label-efficient and robust learning / Recognition and long-tail
AFD-Net: Aggregated Feature Difference Learning for Cross-Spectral Image Patch Matching
Dou Quan, Xuefeng Liang, Shuang Wang, Shaowei Wei, Yanfeng Li, Ning Huyan, 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 |
|---|---|---|---|
| add aggregated multi-level feature differences (AD(5,4), AD(5,4,3), etc.) over a plain Siamese network AD(0) (ablation) | AFD-Net · VIS-NIR | FPR95 (%): mean improvements 32.20% to 38.98% | Improved |
| aggregate feature differences down to levels 2 and 1 beyond AD(5,4,3) (ablation) | AFD-Net · VIS-NIR | FPR95 (%) | Mixed |
| add instance normalization on top of batch normalization in shallow layers (IBN vs BN) (ablation) | AFD-Net · VIS-NIR (train Country, test 8 categories, mean) | FPR95 (%): 1.47 -> 1.08 | Improved |
| add batch normalization (BN-LMCL vs no normalization) (ablation) | AFD-Net · VIS-NIR (train Country, test 8 categories, mean) | FPR95 (%): 2.09 -> 1.47 | Improved |
| replace Softmax loss with large margin cosine loss (LMCL) (ablation) | AFD-Net · VIS-NIR (train Country, test 8 categories, mean) | FPR95 (%): 1.82 -> 1.08 | Improved |
| raise LMCL margin m from 0.15 to 0.25 at s=20 (ablation) | AFD-Net · VIS-NIR (train Country, test Field) | FPR95 (%): 4.12 -> 3.47 | Improved |
| raise LMCL margin m from 0.25 to 0.35 at s=20 (ablation) | AFD-Net · VIS-NIR (train Country, test Field) | FPR95 (%): 3.47 -> 4.86 | Hurt |
| use AFD-Net instead of HardNet on cross-spectral matching | AFD-Net DA · VIS-NIR (train Country, test 8 categories, mean) | FPR95 (%): 2.80 -> 1.08 | Improved |
| use AFD-Net instead of HardNet on single-spectral Brown dataset | AFD-Net DA · Multi-view stereo correspondence (Brown, mean of 6 splits) | FPR95 (%): 1.51 -> 1.42 | Improved |
Closest papers by meaning
- Understanding Dataset Distillation via Spectral Filtering (ICLR 2026 · similarity 0.85)
- Unsupervised Cross-Spectral Stereo Matching by Learning to Synthesize (AAAI 2019 · similarity 0.85)
- Better and Faster: Exponential Loss for Image Patch Matching (ICCV 2019 · similarity 0.85)
- HAP: Harmonized Amplitude Perturbation for Cross-Domain Few-Shot Learning (AAAI 2026 · similarity 0.85)
- From Feature Learning to Spectral Basis Learning: A Unifying and Flexible Framework for Efficient and Robust Shape Matching (CVPR 2026 · similarity 0.85)
- Towards an Explainable Comparison and Alignment of Feature Embeddings (ICML 2025 · 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.
