Research map / Architectures and efficiency / CNN design
Spatially-Adaptive Filter Units for 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 |
|---|---|---|---|
| replace regular convolution filters with learned displaced aggregation units | AlexNet · ILSVRC 2012 validation | top-1 accuracy (%): 56.99 -> 56.89 | No clear effect |
| replace regular convolution filters with learned displaced aggregation units | AlexNet · ILSVRC 2012 validation | convolutional-layer parameters: 3.7 mio -> 2.3 mio | Improved |
| use learned DAU filters instead of dilated convolution | AlexNet segmentation network · PASCAL VOC 2011 validation | mean IU: 45.57 -> 47.22 | Improved |
| vary Gaussian aggregation variance from 0.3^2 to 0.8^2 (ablation) | DAU-ConvNet · CIFAR10 | classification performance: 82.9, 83.4, 83.8, 83.6, 82.9, 82.8 | No clear effect |
| increase DAUs per filter from 1 to 6 (ablation) | DAU-ConvNet · CIFAR10 | classification performance: 82.9 -> 84.1 | Improved |
| use 1–2 DAUs per filter rather than more units (ablation) | DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 57.3; 56.9; 56.4 | No clear effect |
| remove units below relative amplification threshold 0.01 (ablation) | Large DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 57.3 -> 57.3 | No clear effect |
| remove units below relative amplification threshold 0.02 (ablation) | Large DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 57.3 -> 57.1 | No clear effect |
| remove units below relative amplification threshold 0.05, 0.10, or 0.25 (ablation) | Large DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 57.3 -> 40.1; 28.3; 0.2 | Hurt |
| remove units below relative amplification threshold 0.01 or 0.02 (ablation) | Medium DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 56.9 -> 56.8; 56.7 | No clear effect |
| remove units below relative amplification threshold 0.05, 0.10, or 0.25 (ablation) | Medium DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 56.9 -> 54.8; 47.4; 1.9 | Hurt |
| remove units below relative amplification threshold 0.01 or 0.02 (ablation) | Small DAU-ConvNet · ILSVRC 2012 validation | top-1 accuracy (%): 56.4 -> 56.4 | No clear effect |
Closest papers by meaning
- Spatially-Adaptive Filter Units for Compact and Efficient Deep Neural Networks (IJCV 2020 · similarity 0.96)
- Integrating Multiple Receptive Fields Through Grouped Active Convolution (TPAMI 2020 · similarity 0.86)
- Adaptive Convolutions With Per-Pixel Dynamic Filter Atom (ICCV 2021 · similarity 0.85)
- Understanding the Effective Receptive Field in Deep Convolutional Neural Networks (NeurIPS 2016 · similarity 0.85)
- 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.
