Research map / Detection and segmentation / CNN features
Learnable Polyphase Sampling for Shift Invariant and Equivariant Convolutional Networks
Renan A. Rojas Gomez, Teck-Yian Lim, Alex Schwing, Minh N. Do, Raymond A. Yeh
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 standard downsampling with LPS | ResNet-18 · CIFAR10, circular shift | top-1 accuracy: 91.44 ± 0.33 -> 94.45 ± 0.05 | Improved |
| Replace APS with LPS | ResNet-18 · CIFAR10, circular shift | top-1 accuracy: 94.07 ± 0.28 -> 94.45 ± 0.05 | Improved |
| Use LPS instead of LPF with Tri-3 filtering | ResNet-50 · ImageNet, standard shift | S-Cons.: 90.91 -> 91.43 | Improved |
| Use LPS instead of DDAC* | DeepLabV3+ with ResNet-101 · PASCAL VOC, standard shift | mIoU: 80.31 -> 80.43 | Improved |
| Add LPU to the LPD backbone (ablation) | DeepLabV3+ with ResNet-18 · PASCAL VOC, circular shift | mASCC: 96.23 -> 100 | Improved |
| Train with Gumbel-Softmax sampling rather than standard Softmax (ablation) | LPS ResNet-18 with Tri-3 · CIFAR10, circular shift | top-1 accuracy: 93.22 ± 0.13 -> 94.8 ± 0.14 | Improved |
| Replace LPD layers with APS (ablation) | ResNet-101 · ImageNet validation | top-1 accuracy: 78.8% -> 0.1% | Hurt |
| Combine LPS with Rect-2 anti-aliasing rather than LPS without anti-aliasing | ResNet-18 · CIFAR10, circular shift | top-1 accuracy: 94.45 ± 0.05 -> 94.69 ± 0.06 | Improved |
Closest papers by meaning
- Truly Shift-Invariant Convolutional Neural Networks (CVPR 2021 · similarity 0.90)
- Making Convolutional Networks Shift-Invariant Again (ICML 2019 · similarity 0.85)
- How Shift Equivariance Impacts Metric Learning for Instance Segmentation (ICCV 2021 · similarity 0.85)
- Empowering Networks With Scale and Rotation Equivariance Using A Similarity Convolution (ICLR 2023 · similarity 0.84)
- Equivariance versus Augmentation for Spherical Images (ICML 2022 · similarity 0.84)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 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.
