Research map / Architectures and efficiency / CNN design
Spatially-Adaptive Filter Units for Compact and Efficient Deep Neural Networks
Domen Tabernik, Matej Kristan, Aleš Leonardis
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 |
|---|---|---|---|
| learn standard deviation per unit instead of fixing 0.5 (ablation) | DAU-ConvNet · CIFAR10 (shallow 3-layer DAU-ConvNet) | accuracy (%): 83.8 -> 84.25 | No clear effect |
| vary fixed standard deviation 0.3 to 0.8 (ablation) | DAU-ConvNet · CIFAR10 (shallow 3-layer DAU-ConvNet) | accuracy (%): 82.8 to 83.8 | No clear effect |
| prune units with relative threshold 0.01 to 0.02 (ablation) | Large DAU-AlexNet · ILSVRC 2012 val | top-1 accuracy (%): 57.3 -> 57.1 | No clear effect |
| prune units with relative threshold 0.05 (ablation) | Large DAU-AlexNet · ILSVRC 2012 val | top-1 accuracy (%): 57.3 -> 40.1 | Hurt |
| prune units with relative threshold 0.05 (ablation) | Small DAU-AlexNet · ILSVRC 2012 val | top-1 accuracy (%): 56.4 -> 55.4 | Hurt |
| prune units with relative threshold 0.25 (ablation) | Medium DAU-AlexNet · ILSVRC 2012 val | top-1 accuracy (%): 56.9 -> 1.9 | Hurt |
| replace conv filters by DAUs (Medium) in AlexNet | AlexNet · ILSVRC 2012 val | top-1 accuracy (%): 56.99 -> 56.89 with 3.7 million -> 2.3 million params | No clear effect |
| replace 3x3 convs by two-unit DAUs | ResNet101 · ILSVRC 2012 val | top-1 accuracy (%): 75.39 -> 74.89 | Hurt |
| replace dilated convolution with DAUs | AlexNet · PASCAL VOC 2011 val | mIoU (%): 45.57 -> 47.22 | Improved |
| replace convolutions by DAUs | ResNet101 (OS16) · Cityscapes val | mIoU (%): 68.6 -> 72.8 | Improved |
| add image-pooling features (ablation) | ResNet101 · Cityscapes val | mIoU (%): 68.6 -> 72.7 | Improved |
| replace ASPP with single DAU-6U layer | DeepLabv3+ ResNet101 · Cityscapes val | mIoU (%): 75.6 -> 75.5 | No clear effect |
Closest papers by meaning
- Spatially-Adaptive Filter Units for Deep Neural Networks (CVPR 2018 · similarity 0.96)
- Learning Versatile Convolution Filters for Efficient Visual Recognition (TPAMI 2021 · similarity 0.88)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.87)
- Learning Versatile Filters for Efficient Convolutional Neural Networks (NeurIPS 2018 · similarity 0.86)
- Receptive Fields As Experts in Convolutional Neural Architectures (ICML 2024 · similarity 0.85)
- The Master Key Filters Hypothesis: Deep Filters Are General (AAAI 2025 · similarity 0.84)
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.
