Research map / Detection and segmentation / CNN features
Structured Binary Neural Networks for Image Recognition
Bohan Zhuang, Chunhua Shen, Mingkui Tan, Peng Chen, Lingqiao Liu, Ian D. Reid
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 learned soft gates (ablation) | GroupNet-A (ResNet-34) · ImageNet val | Top-1 accuracy: 66.4 -> 68.7 | Improved |
| add learned soft gates (ablation) | GroupNet-A (ResNet-18) · ImageNet val | Top-1 accuracy: 64.1 -> 65.2 | Improved |
| add learned soft gates (ablation) | GroupNet-A (ResNet-50) · ImageNet val | Top-1 accuracy: 67.7 -> 69.8 | Improved |
| keep 1x1 downsampling skip connections at 8-bit (GroupNet-B vs A) | GroupNet · ImageNet val (ResNet-18) | Top-1 accuracy: 65.2 -> 67.3 | Improved |
| replace soft routing with dynamic hard gating (GroupNet-C vs B) | GroupNet · ImageNet val (ResNet-18) | Top-1 accuracy: 67.3 -> 68.2 | Improved |
| group decomposition instead of width multiplier 2 (ablation) | GroupNet-A · ImageNet val (ResNet-18) | Top-1 accuracy: 63.5 -> 65.2 | Improved |
| group-wise (block) decomposition instead of layer-wise (ablation) | GroupNet-A · ImageNet val (ResNet-18) | Top-1 accuracy: 57.8 -> 65.2 | Improved |
| one block per base instead of whole network per base (ablation) | GroupNet · ImageNet val (ResNet-18) | Top-1 accuracy: 59.8 -> 65.2 | Improved |
| increase number of bases from 1 to 5 (ablation) | GroupNet-A · ImageNet val (ResNet-18) | Top-1 accuracy: 58.1 -> 66.0 | Improved |
| use diverse dilation rates across binary bases (BPAC) | GroupNet-C · PASCAL VOC 2012 val (ResNet-18 FCN-8s-C5) | mIOU: 61.5 -> 66.2 | Improved |
| binarize ASPP activations to {0,1} instead of {-1,1} (ablation) | binary DeepLabv3 · PASCAL VOC 2012 val (ResNet-18) | mIOU: 51.6 -> 64.0 | Improved |
| binarize the ASPP module in addition to the backbone (ablation) | binary DeepLabv3 · PASCAL VOC 2012 val (ResNet-34) | mIOU: 71.9 -> 68.0 | Hurt |
Closest papers by meaning
- Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation (CVPR 2019 · similarity 0.96)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.88)
- Binarizing MobileNet via Evolution-Based Searching (CVPR 2020 · similarity 0.88)
- Sub-Bit Neural Networks: Learning To Compress and Accelerate Binary Neural Networks (ICCV 2021 · similarity 0.88)
- Training binary neural networks with real-to-binary convolutions (ICLR 2020 · similarity 0.87)
- Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · similarity 0.87)
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.
