rgeirhos/Stylized-ImageNet

What it solves

It addresses the tendency of Convolutional Neural Networks (CNNs) to rely too heavily on local textures rather than global shapes when identifying objects. By providing a way to create a stylized version of ImageNet, it allows researchers to train models that are more biased toward shape, which can improve overall accuracy and robustness.

How it works

The project uses the AdaIN (Adaptive Instance Normalization) style transfer approach to apply the styles of various paintings to standard ImageNet images. This process distorts local textures while keeping the global object shapes intact, effectively forcing a model to learn shape-based features during training.

Who it’s for

It is designed for AI researchers and developers working on computer vision, specifically those studying model generalization, robustness, and the bias between texture and shape in deep learning models.

Highlights

• Provides a complete pipeline and script to generate the Stylized-ImageNet dataset.
• Includes a Docker image to simplify the installation of required libraries.
• Compatible as a drop-in replacement for the standard ImageNet dataset during training.
• Based on the AdaIN arbitrary style transfer method.