Stylized-ImageNet

Why we included this project

This repository is the reference implementation behind a widely cited result in computer vision: ImageNet-trained CNNs rely far more on texture than on shape, and steering them toward shape makes them more accurate and robust. If you are training a vision model and want to know whether it is actually recognizing object geometry instead of matching local patterns, the code here gives you a direct way to find out. It builds Stylized-ImageNet by transferring paintings onto ImageNet images, which heavily distorts local textures while leaving global shapes largely intact, so you can run your network on the stylized set and see whether it still identifies the object. The setup expects you to supply both the ImageNet images and the painter-by-numbers style set, so it suits researchers and engineers who already work with those resources. If you want a lighter path, the README points to the bethgelab model-vs-human toolbox, which packages stylized data, human comparison data, and turnkey PyTorch/TensorFlow evaluation.