Research map / Label-efficient and robust learning / Recognition and long-tail
Cross-Domain Image Matching with Deep Feature Maps
Bailey Kong, James Steven Supančič, Deva Ramanan, Charless C. Fowlkes
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 |
|---|---|---|---|
| use per-channel normalization (MCNCC) instead of unnormalized correlation (ablation) | NCC on ResNet-50 res2bx · Partial print matching | mAP: 0.15 -> 0.77 | Improved |
| use per-channel statistics instead of whole-volume statistics (standardize) (ablation) | ResNet-50 res2bx · Partial print matching | mAP: 0.32 -> 0.77 | Improved |
| use per-channel MCNCC instead of 3D standardization (ablation) | GoogleNet conv2x · Partial print matching | mAP: 0.68 -> 0.81 | Improved |
| use per-channel MCNCC instead of 3D standardization (ablation) | DeepVGG-16 · Partial print matching | mAP: 0.73 -> 0.76 | Improved |
| add per-channel standardization by sigma_c on top of centering (ablation) | ResNet-50 res2bx · Partial print matching | mAP: 0.55 -> 0.77 | Improved |
| add global PCA decorrelation of channels before MCNCC (ablation) | MCNCC on ResNet features · Partial print matching | mAP | Hurt |
| add learned per-channel importance weights and CCA projections (ablation) | MCNCC · FID-300 whole prints | top-1% correct matches: 72.67 -> 79.67 | Improved |
| add learned CCA projections and per-channel weights (ablation) | MCNCC · FID-300 occlusion study, all prints | top-1% correct matches: 72.7 -> 79.0 | Improved |
| add joint fine-tuning of projections and weights (ablation) | CCA [mu_c,sigma_c.W_c] · Israeli prints (50/50 split) | recall | No clear effect |
| add learned CCA and weights to MCNCC for 1/4 size prints (ablation) | MCNCC · FID-300 occlusion study, 1/4 size | top-10% correct matches: 84.1 -> 79.4 | Hurt |
| add learned parameters to MCNCC (ablation) | MCNCC · CMP Facade label retrieval | top-1 accuracy: 39.2% -> 96.4% | Improved |
| use MCNCC instead of cosine similarity (ablation) | NCC · CMP Facade label retrieval | top-1 accuracy: 18.3% -> 39.2% | Improved |
Closest papers by meaning
- CriSp: Leveraging Tread Depth Maps for Enhanced Crime-Scene Shoeprint Matching (ECCV 2024 · similarity 0.88)
- Cross-Domain Visual Matching via Generalized Similarity Measure and Feature Learning (TPAMI 2016 · similarity 0.87)
- Feature Representation Learning for Unsupervised Cross-Domain Image Retrieval (ECCV 2022 · similarity 0.84)
- Image Matching from Handcrafted to Deep Features: A Survey (IJCV 2020 · similarity 0.84)
- OmniGlue: Generalizable Feature Matching with Foundation Model Guidance (CVPR 2024 · similarity 0.83)
- ARNet: Self-Supervised FG-SBIR with Unified Sample Feature Alignment and Multi-Scale Token Recycling (AAAI 2025 · similarity 0.82)
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.
