Research map / Architectures and efficiency / CNN design
When Naive Bayes Nearest Neighbors Meet Convolutional Neural Networks
Ilja Kuzborskij, Fabio Maria Carlucci, Barbara Caputo
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 patches from multiple scales instead of using only one patch scale (ablation) | CNN-NBNN · Scene 15, UIUC Sports, and MIT Indoor | accuracy | Improved |
| Use larger patches rather than only smaller patches (ablation) | CNN-NBNN · Scene recognition | accuracy | Improved |
| Increase sampling density from sparse to dense (ablation) | CNN-NBNN · Scene recognition | accuracy | No clear effect |
| Use 16px patches (ablation) | CNN-NBNN · Scene recognition | accuracy | Mixed |
| Use CNN features before ReLU rather than after ReLU (ablation) | CNN-NBNN · Scene recognition | accuracy | Improved |
| Replace NBNL with scalable NBNL | NBNL · Scene 15, Sports 8, and MIT Indoor | accuracy: 91.5 -> 91.638; 94.29 -> 95.268; 72.592 -> 72.786 | No clear effect |
| Replace NBNL with scalable NBNL | NBNL · Scene 15, Sports 8, and MIT Indoor | training time: 5729.2 -> 210.3; 1024.4 -> 63.564; 9690 -> 304.26 | Improved |
| Use CNN-sNBNL rather than a linear SVM on Hybrid CNN features | SUN-397 scene recognition · SUN-397 | accuracy: 53.86 ± 0.21% -> 55.8 ± 0.29% | Improved |
| Use CNN-sNBNL rather than a linear SVM on Places-205 CNN features | SUN-397 scene recognition · SUN-397 | accuracy: 54.32 ± 0.14% -> 55.8 ± 0.29% | Improved |
| Use CNN-sNBNL rather than CNN-NBNL | CNN-NBNL · UIUC Sports | accuracy: 95.29 ± 0.61 -> 95.28 ± 0.68 | No clear effect |
| Use CNN-sNBNL rather than CNN-NBNL | CNN-NBNL · MIT Indoor | accuracy: 73 ± 0.36 -> 72.79 ± 0.73 | No clear effect |
Closest papers by meaning
- Improving Generalization via Scalable Neighborhood Component Analysis (ECCV 2018 · similarity 0.86)
- Visual Recognition with Deep Nearest Centroids (ICLR 2023 · similarity 0.85)
- An Empirical Evaluation of Current Convolutional Architectures' Ability to Manage Nuisance Location and Scale Variability (CVPR 2016 · similarity 0.85)
- Is Second-Order Information Helpful for Large-Scale Visual Recognition? (ICCV 2017 · similarity 0.85)
- Network of Experts for Large-Scale Image Categorization (ECCV 2016 · similarity 0.85)
- Object Detection Networks on Convolutional Feature Maps (TPAMI 2016 · 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.
