Research map / Architectures and efficiency / CNN design
Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation
Check 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 |
|---|---|---|---|
| Use group-wise rather than layer-wise binary decomposition (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 57.6 -> 64.8 | Helped |
| Learn soft connections between blocks rather than use fixed hard group connections (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy | Helped |
| Add more bases, from 1 to 3 (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 55.6 -> 62.5 | Helped |
| Add more bases, from 3 to 5 (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 62.5 -> 64.8 | Helped |
| Add shortcut bypasses around binary convolutions (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 64.8 -> 67.0 | Helped |
| Replace one-block bases with two-block bases (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 63.0 -> 62.2 | Hurt |
| Make each base a whole binary network rather than use learned group connections (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 64.8 -> 59.2 | Hurt |
| Add BPAC with varied dilation rates across binary bases (ablation) | Group-Net · PASCAL VOC 2012 val, ResNet-18 FCN-32s | mIOU: 60.5 -> 63.8 | Helped |
| Add shortcut bypasses to Group-Net + BPAC (ablation) | Group-Net + BPAC · PASCAL VOC 2012 val, ResNet-18 FCN-32s | mIOU: 63.8 -> 65.1 | Helped |
| Use multi-dilation branches in a full-precision backbone (ablation) | Full-precision FCN · PASCAL VOC 2012 val, ResNet-18 FCN-32s | mIOU: 64.9 -> 67.6 | Helped |
| Replace binary activations with 2-bit activations while retaining binary weights (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 64.8 -> 69.6 | Helped |
| Replace binary activations with 32-bit activations while retaining binary weights (ablation) | Group-Net · ImageNet val, ResNet-18 | Top-1 accuracy (%): 64.8 -> 70.4 | Helped |
Closest papers by meaning
- Binarizing MobileNet via Evolution-Based Searching (CVPR 2020 · similarity 0.87)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.86)
- GroupViT: Semantic Segmentation Emerges From Text Supervision (CVPR 2022 · similarity 0.86)
- Group-Wise Semantic Mining for Weakly Supervised Semantic Segmentation (AAAI 2021 · similarity 0.86)
- Quantized Convolutional Neural Networks for Mobile Devices (CVPR 2016 · similarity 0.86)
- Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · 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.