# Mind the Pad -- CNNs can Develop Blind Spots

arXiv: https://arxiv.org/abs/2010.02178 · HTML: https://mlsanity.com/paper/2010.02178/

## Key facts

- Cited by 41 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 5 Oct 2020.

## Built on by (most cited first, 30 of 41)

- [Simple Baselines for Image Restoration](https://mlsanity.com/paper/2204.04676/) (2022)
- [Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer](https://mlsanity.com/paper/2204.03638/) (2022)
- [Feature-Realistic Neural Fusion for Real-Time, Open Set Scene Understanding](https://mlsanity.com/paper/2210.03043/) (2022)
- [Stabilizing Off-Policy Deep Reinforcement Learning from Pixels](https://mlsanity.com/paper/2207.00986/) (2022)
- [Deep Image Fingerprint: Towards Low Budget Synthetic Image Detection and Model Lineage Analysis](https://mlsanity.com/paper/2303.10762/) (2023)
- [FoveaTer: Foveated Transformer for Image Classification](https://mlsanity.com/paper/2105.14173/) (2021)
- [Position, Padding and Predictions: A Deeper Look at Position Information in CNNs](https://mlsanity.com/paper/2101.12322/) (2021)
- [High-fidelity GAN Inversion with Padding Space](https://mlsanity.com/paper/2203.11105/) (2022)
- [GANSeg: Learning to Segment by Unsupervised Hierarchical Image Generation](https://mlsanity.com/paper/2112.01036/) (2021)
- [Toward Spatially Unbiased Generative Models](https://mlsanity.com/paper/2108.01285/) (2021)
- [Shift Invariance Can Reduce Adversarial Robustness](https://mlsanity.com/paper/2103.02695/) (2021)
- [Effects of boundary conditions in fully convolutional networks for learning spatio-temporal dynamics](https://mlsanity.com/paper/2106.11160/) (2021)
- [Classification of Long Sequential Data using Circular Dilated Convolutional Neural Networks](https://mlsanity.com/paper/2201.02143/) (2022)
- [Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs](https://mlsanity.com/paper/2108.07884/) (2021)
- [Cylin-Painting: Seamless 360textdegree Panoramic Image Outpainting and Beyond](https://mlsanity.com/paper/2204.08563/) (2022)
- [Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks](https://mlsanity.com/paper/2202.00838/) (2022)
- [RecRecNet: Rectangling Rectified Wide-Angle Images by Thin-Plate Spline Model and DoF-based Curriculum Learning](https://mlsanity.com/paper/2301.01661/) (2023)
- [Foundation Models and Transformers for Anomaly Detection: A Survey](https://mlsanity.com/paper/2507.15905/) (2025)
- [How Shift Equivariance Impacts Metric Learning for Instance Segmentation](https://mlsanity.com/paper/2101.05846/) (2021)
- [Prescriptive and Descriptive Approaches to Machine-Learning Transparency](https://mlsanity.com/paper/2204.13582/) (2022)
- [Zone Evaluation: Revealing Spatial Bias in Object Detection](https://mlsanity.com/paper/2310.13215/) (2023)
- [Hierarchical Disentanglement-Alignment Network for Robust SAR Vehicle Recognition](https://mlsanity.com/paper/2304.03550/) (2023)
- [NeRD: Neural Representation of Distribution for Medical Image Segmentation](https://mlsanity.com/paper/2103.04020/) (2021)
- [LatentKeypointGAN: Controlling Images via Latent Keypoints](https://mlsanity.com/paper/2103.15812/) (2021)
- [Unveiling The Mask of Position-Information Pattern Through the Mist of Image Features](https://mlsanity.com/paper/2206.01202/) (2022)
- [Generalization to translation shifts: a study in architectures and augmentations](https://mlsanity.com/paper/2207.02349/) (2022)
- [Improved Padding in CNNs for Quantitative Susceptibility Mapping](https://mlsanity.com/paper/2106.15331/) (2021)
- [Seamless Satellite-image Synthesis](https://mlsanity.com/paper/2111.03384/) (2021)
- [Auditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions](https://mlsanity.com/paper/2602.02560/) (2026)
- [AdaSpot: Spend Resolution Where It Matters for Precise Event Spotting](https://mlsanity.com/paper/2602.22073/) (2026)

## Abstract (arXiv)

We show how feature maps in convolutional networks are susceptible to spatial bias. Due to a combination of architectural choices, the activation at certain locations is systematically elevated or weakened. The major source of this bias is the padding mechanism. Depending on several aspects of convolution arithmetic, this mechanism can apply the padding unevenly, leading to asymmetries in the learned weights. We demonstrate how such bias can be detrimental to certain tasks such as small object detection: the activation is suppressed if the stimulus lies in the impacted area, leading to blind spots and misdetection. We propose solutions to mitigate spatial bias and demonstrate how they can improve model accuracy.