Research map / Architectures and efficiency / CNN design
An intriguing failing of convolutional neural networks and the CoordConv solution
Read the paperCheck 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 |
|---|---|---|---|
| Add coordinate channels in the first layer | Convolutional coordinate classifier · Supervised Coordinate Classification, uniform split | test accuracy: 86% -> 100% | Helped |
| Add coordinate channels in the first layer | Convolutional coordinate classifier · Supervised Coordinate Classification, quadrant split | test accuracy | Helped |
| Add a CoordConv layer | Convolutional coordinate regressor · Supervised Coordinate Regression, uniform and quadrant splits | generalization: around five pixels error on average | Helped |
| Replace the first layer with CoordConv | Convolutional renderer · Supervised Rendering, uniform and quadrant splits | test IOU: about IOU 0.83 -> perfect test IOU | Helped |
| Add a 1×1 CoordConv layer at the beginning | ResNet-50 · ImageNet | Top-5 accuracy: 0.92712 -> 0.927564; p=.11 | No clear effect |
| Add CoordConv to the detection model | Faster R-CNN · MNIST digits detection, test set | Test IOU-average: 0.1508 -> 0.1868; 23.87 | Helped |
| Use CoordConv in the generator | GAN · Colored-shape generation | red-object position entropy: 3.13 -> 3.30 | Helped |
| Use CoordConv in the generator | GAN · Colored-shape generation | relative-position entropy: 2.81 -> 2.62 | Hurt |
| Add a CoordConv layer to the actor network | A2C · Atari games | score: 6 of 9 games improve; 2 perform similarly; 1 is slightly worse | Mixed |
| Add CoordConv to the model | Distributed Prioritized Experience Replay (Ape-X) · Atari games | performance | No clear effect |
| Use CoordConv in both generator and discriminator instead of only in the generator (ablation) | Colored-shape GAN · Colored-shape generation | sample quality | No clear effect |
| Use CoordConv in the GAN generator (ablation) | Colored-shape GAN · Colored-shape generation | relative locations between objects: GAN 2.81; CoordConv GAN 2.62 | Hurt |
Closest papers by meaning
- Patch Reordering: A NovelWay to Achieve Rotation and Translation Invariance in Convolutional Neural Networks (AAAI 2017 · similarity 0.87)
- Warped Convolutions: Efficient Invariance to Spatial Transformations (ICML 2017 · similarity 0.87)
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.86)
- CoordX: Accelerating Implicit Neural Representation with a Split MLP Architecture (ICLR 2022 · similarity 0.86)
- MogaNet: Multi-order Gated Aggregation Network (ICLR 2024 · similarity 0.84)
- REOrdering Patches Improves Vision Models (NeurIPS 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.