Quantifying and Inducing Shape Bias in CNNs via Max-Pool Dilation

TL;DR

This paper introduces a data-driven metric to quantify shape versus texture bias in datasets and proposes a max-pool dilation method to induce shape bias in CNNs, improving performance on shape-dominant data especially with limited training data.

Contribution

The paper presents a novel metric for dataset shape-texture analysis and a max-pool dilation technique to enhance shape bias in CNNs without retraining all weights.

Findings

Improves accuracy on shape-dominant datasets.

Effective in low-data regimes with minimal training.

Maintains convolutional weights while inducing shape bias.

Abstract

Convolutional Neural Networks (CNNs) exhibit a well-known texture bias, prioritizing local patterns over global shapes - a tendency inherent to their convolutional architecture. While this bias is beneficial for texture-rich natural images, it often degrades performance on shape-dominant data such as illustrations and sketches. Although prior work has proposed shape-biased models to mitigate this issue, these approaches lack a quantitative metric for identifying which datasets would actually benefit from such modifications. To address this limitation, we propose a data-driven metric that quantifies the shape-texture balance within a dataset by computing the Structural Similarity Index (SSIM) between an image's luminance (Y) channel and its L0-smoothed counterpart. Building on this metric, we introduce a computationally efficient adaptation method that promotes shape bias by modifying the dilation of max-pooling operations while keeping convolutional weights frozen. Experimental results demonstrate consistent accuracy improvements on shape-dominant datasets, particularly in low-data regimes where full fine-tuning is impractical, requiring training only the final classification layer.