Research map / Architectures and efficiency / CNN design
Data-Driven Neuron Allocation for Scale Aggregation Networks
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 |
|---|---|---|---|
| Replace ResNet 3×3 convolutions with learnable scale aggregation blocks | ResNet on ImageNet · ImageNet validation | top-1 error: 24.02 -> 22.20; 22.09 -> 20.97; 21.58 -> 20.62 | Helped |
| Replace ResNet with ScaleNet as Faster R-CNN backbone | Faster R-CNN on COCO · COCO mini-validation, 600/1000 and 800/1200 | mmAP: 31.7 -> 36.2; 34.1 -> 37.3; 32.6 -> 37.2; 35.9 -> 39.5 | Helped |
| Replace ResNet bottleneck 3×3 convolutions with scale aggregation blocks | ResNet on CIFAR-100 · CIFAR-100 | top-1 error: 26.88 -> 24.60; 26.19 -> 23.83; 24.54 -> 22.77 | Helped |
| Use data-driven neuron allocation rather than even scale allocation (ablation) | ScaleNet · ImageNet validation | top-1 error: 22.76 -> 22.20 | Helped |
| Use data-driven rather than even neuron allocation across scales (ablation) | ScaleNet · CIFAR-100 and ImageNet, scale counts L=2 to 5 | top-1 error | Mixed |
| Use max pooling rather than stride-2 3×3 convolution for downsampling (ablation) | ScaleNet-56 · CIFAR-100, L=2 | top-1 error: 26.19 -> 24.48 | Helped |
| Use max pooling rather than dilated stride-2 3×3 convolution for downsampling (ablation) | ScaleNet-56 · CIFAR-100, L=2 | top-1 error: 25.42 -> 24.48 | Helped |
| Use max pooling rather than average pooling for downsampling (ablation) | ScaleNet-56 · CIFAR-100, L=2 | top-1 error: 24.58 -> 24.48 | Helped |
| Use the light ScaleNet-50 variant instead of ResNet-50 | ScaleNet-50 · ImageNet validation, single crop | GFLOPs: 4.1 -> 2.9 | Helped |
Closest papers by meaning
- ScaleNet: A Shallow Architecture for Scale Estimation (CVPR 2022 · similarity 0.89)
- ScaleNet - Improve CNNs through Recursively Rescaling Objects (AAAI 2020 · similarity 0.89)
- Deep Layer Aggregation (CVPR 2018 · similarity 0.89)
- Revisiting ResNets: Improved Training and Scaling Strategies (NeurIPS 2021 · similarity 0.88)
- Scene Recognition With CNNs: Objects, Scales and Dataset Bias (CVPR 2016 · similarity 0.88)
- LSNet: See Large, Focus Small (CVPR 2025 · 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.