Research map / Architectures and efficiency / CNN design
Shape Adaptor: A Learnable Resizing Module
Shikun Liu, Zhe Lin, Yilin Wang, Jianming Zhang, Federico Perazzi, Edward Johns
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 human-designed resizing layers with shape adaptors | MobileNetv2 · Birds | accuracy (%): 60.37 -> 68.53 | Improved |
| replace human-designed resizing layers with shape adaptors | MobileNetv2 · Aircraft | accuracy (%): 77.64 -> 83.00 | Improved |
| replace human-designed resizing layers with shape adaptors | ResNet-50 · CIFAR-100 | accuracy (%): 78.53 -> 80.29 | Improved |
| replace human-designed resizing layers with shape adaptors | ResNet-50 · CIFAR-10 | accuracy (%): 95.50 -> 95.48 | No clear effect |
| replace human-designed resizing layers with shape adaptors | VGG-16 · ImageNet | accuracy (%): 73.92 -> 73.53 | Hurt |
| replace human-designed resizing layers with shape adaptors | ResNet-50 · ImageNet | accuracy (%): 77.18 -> 78.74 | Improved |
| use learned shape during training instead of retraining with fixed final shape (Fix Final) (ablation) | Shape Adaptor VGG-16 · CIFAR-100 | accuracy (%): 78.62 -> 79.16 | Improved |
| use learned shape instead of fixed largest shape (Fix Large) (ablation) | Shape Adaptor VGG-16 · Aircraft | accuracy (%): 84.49 -> 86.95 | Improved |
| use 6 shape adaptors per heuristic instead of 5 (ablation) | VGG-16 · Aircraft | accuracy (%): 84.80 -> 86.95 | Improved |
| use 4 shape adaptors per heuristic instead of 8 (ablation) | VGG-16 · CIFAR-100 | accuracy (%): 78.16 -> 79.16 | Improved |
| initialise s(alpha) at 0.52 instead of heuristic 0.58 (ablation) | Shape Adaptor VGG-16 · Aircraft | accuracy (%): 83.90 -> 86.95 | Improved |
| apply AutoSC shape compression to MobileNetv2 | MobileNetv2 · Aircraft | accuracy (%): 77.64 -> 78.95 | Improved |
Closest papers by meaning
- Learning To Resize Images for Computer Vision Tasks (ICCV 2021 · similarity 0.86)
- NeuForm: Adaptive Overfitting for Neural Shape Editing (NeurIPS 2022 · similarity 0.85)
- Straight to Shapes: Real-Time Detection of Encoded Shapes (CVPR 2017 · similarity 0.85)
- Aggregated Residual Transformations for Deep Neural Networks (CVPR 2017 · similarity 0.84)
- Revisiting ResNets: Improved Training and Scaling Strategies (NeurIPS 2021 · similarity 0.84)
- $\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation (ICLR 2024 · similarity 0.83)
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.
