Research map / Architectures and efficiency / CNN design
Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet
Wieland Brendel, Matthias Bethge
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 |
|---|---|---|---|
| Limit classification to local features within 33 × 33 pixel patches and average patch evidence linearly | BagNet-33 · ImageNet | top-5 accuracy: 87.6% | Improved |
| Limit classification to local features within 17 × 17 pixel patches and average patch evidence linearly | BagNet-17 · ImageNet | top-5 accuracy: 80.5% | Improved |
| Use BagNet inference instead of ResNet-50 inference | BagNet-9, BagNet-17, BagNet-33 · ImageNet, 3 × 224 × 224 images, batch size 64 | inference throughput: around 155 images/s vs 570 images/s | Hurt |
| Texturise images while preserving local feature statistics | VGG-16 · ImageNet texturised images | accuracy: 90.1% on clean versus 79.4% on texturised image | Hurt |
| Use BagNet-17 heatmaps to select patches for masking | VGG-16 · ImageNet, masking 100 highest-attribution 8 × 8 patches | average probability of leading class: 0.151 | Improved |
| Use BagNet-17 heatmaps to select patches for masking | ResNet-50 · ImageNet, masking 100 highest-attribution 8 × 8 patches | average probability of leading class: 0.263 | Improved |
| Use BagNet-17 heatmaps to select patches for masking | ResNet-152 · ImageNet, masking 100 highest-attribution 8 × 8 patches | average probability of leading class: 0.371 | Improved |
| Use BagNet-17 heatmaps to select patches for masking | DenseNet-169 · ImageNet, masking 100 highest-attribution 8 × 8 patches | average probability of leading class: 0.326 | Improved |
| Replace patch logits below a threshold with the threshold value (ablation) | BagNet-33 · ImageNet | top-5 accuracy | No clear effect |
| Binarize patch heatmaps by thresholding their values (ablation) | BagNet-33 · ImageNet | top-5 accuracy: within 3-4% of the vanilla BagNet performance | No clear effect |
Closest papers by meaning
- Foveated BagNet: Inherent Interpretability Does Not Exclude Global Context (NeurIPS 2026 · similarity 0.88)
- Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition (TPAMI 2017 · similarity 0.86)
- XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks (ECCV 2016 · similarity 0.86)
- Building a Regular Decision Boundary With Deep Networks (CVPR 2017 · similarity 0.86)
- ImageNet-Hard: The Hardest Images Remaining from a Study of the Power of Zoom and Spatial Biases in Image Classification (NeurIPS 2023 · similarity 0.86)
- Learning Efficient Convolutional Networks Through Network Slimming (ICCV 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.
