Research map / Architectures and efficiency / CNN design
Kernelized Subspace Pooling for Deep Local Descriptors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the projection Gaussian distance kernel to subspace pooling and the marginal triplet loss (ablation) | SP descriptor · UBC mean FPR | FPR at 95% recall: 1.38 -> 0.68 | Helped |
| add the projection Gaussian distance kernel to subspace pooling and the marginal triplet loss (ablation) | SP descriptor · HPatches patch verification | mAP: 88.95% -> 89.14% | Helped |
| add the projection Gaussian distance kernel to subspace pooling and the marginal triplet loss (ablation) | SP descriptor · HPatches image matching | mAP: 54.67% -> 57.80% | Helped |
| add the projection Gaussian distance kernel to subspace pooling and the marginal triplet loss (ablation) | SP descriptor · DaLI deformation and illumination | mean accuracy: 83.00% -> 84.81% | Helped |
| use subspace pooling instead of max pooling (ablation) | global CNN pooling for local descriptors · HPatches image matching | mAP: 39.54% -> 54.67% | Helped |
| use subspace pooling instead of average pooling (ablation) | global CNN pooling for local descriptors · HPatches patch retrieval | mAP: 54.61% -> 63.22% | Helped |
| replace the TF-M fully connected layer with subspace pooling (ablation) | TF-M · HPatches image matching | mAP: 32.64% -> 36.07% | Helped |
| replace the TF-M fully connected layer with kernelized subspace pooling (ablation) | TF-M · HPatches patch retrieval | mAP: 39.40% -> 45.68% | Helped |
| use 16 principal components rather than fewer or more components (ablation) | SP descriptor · HPatches | mAP | Mixed |
| learn the Gaussian bandwidth rather than treating it as a fixed parameter (ablation) | KSP training · multiple datasets | descriptor performance | Hurt |
Closest papers by meaning
- Local Features and Visual Words Emerge in Activations (CVPR 2019 · similarity 0.88)
- Learning Deep Features for Discriminative Localization (CVPR 2016 · similarity 0.88)
- Kernel Pooling for Convolutional Neural Networks (CVPR 2017 · similarity 0.87)
- Local Descriptors Optimized for Average Precision (CVPR 2018 · similarity 0.87)
- Explicit Spatial Encoding for Deep Local Descriptors (CVPR 2019 · similarity 0.86)
- L2-Net: Deep Learning of Discriminative Patch Descriptor in Euclidean Space (CVPR 2017 · similarity 0.86)
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.