Research map / Architectures and efficiency / CNN design
Deep Neural Networks with Box Convolutions
Egor Burkov, Victor Lempitsky
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 |
|---|---|---|---|
| replace dilated/bottleneck blocks of ENet with box convolution blocks (BoxENet) (ablation) | ENet · Cityscapes val | IoU-class: 59.4 -> 64.6 | Improved |
| replace dilated/bottleneck blocks of ENet with box convolution blocks (BoxENet) (ablation) | ENet · SUN RGB-D test | mean IoU: 22.9% -> 24.5% | Improved |
| replace blocks of ERFNet with box blocks (BoxERFNet) (ablation) | ERFNet · Cityscapes val | IoU-class: 68.8 -> 69.0 | No clear effect |
| replace blocks of ERFNet with box blocks (BoxERFNet) (ablation) | ERFNet · SUN RGB-D test | mean IoU: 25.3% -> 28.7% | Improved |
| freeze box coordinates at initialization (ablation) | BoxENet · Cityscapes val | IoU-class: 64.6 -> 60.3 | Hurt |
| freeze box coordinates at initialization (ablation) | BoxERFNet · Cityscapes val | IoU-class: 69.0 -> 63.6 | Hurt |
| freeze box coordinates at initialization (ablation) | BoxENet · SUN RGB-D test | mean IoU: 24.5% -> 21.9% | Hurt |
| remove residual blocks instead of replacing them with box blocks (ablation) | ENet · Cityscapes val | IoU-class: 54.1 vs 64.6 (BoxENet) | Hurt |
| remove residual blocks instead of replacing them with box blocks (ablation) | ERFNet · Cityscapes val | IoU-class: 59.8 vs 69.0 (BoxERFNet) | Hurt |
| replace all resolution-preserving blocks with box blocks (BoxOnlyENet) rather than every second (ablation) | BoxENet · Cityscapes val | IoU-class: 64.6 -> 61.0 | Hurt |
| use box convolutions in ERFNet family: computational cost (ablation) | ERFNet · Resource cost | MultAdds (billions): 30.876 -> 16.042 | Improved |
| use box convolutions in ERFNet family: CPU inference (ablation) | ERFNet · Resource cost | CPU time (ms): 1682 -> 822 | Improved |
Closest papers by meaning
- Segmentation-Aware Convolutional Networks Using Local Attention Masks (ICCV 2017 · similarity 0.88)
- Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells (CVPR 2019 · similarity 0.88)
- Fully Convolutional Networks for Semantic Segmentation (TPAMI 2016 · similarity 0.87)
- Box-Driven Class-Wise Region Masking and Filling Rate Guided Loss for Weakly Supervised Semantic Segmentation (CVPR 2019 · similarity 0.86)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.85)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 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.
