Pre-Training on High-Quality Natural Image Data Reduces DCNN Texture Bias

Deep Convolutional Neural Networks (DCNNs) perform increasingly well on visual tasks like object recognition while also closely aligning with human brain activity.
However, model behaviour also differs from human behaviour in important ways. One prominent example of this difference is that DCNNs trained on ImageNet exhibit a texture bias, while humans are consistently biased towards object shape. Previous work suggests DCNN shape bias can be increased by training on purposely designed stimuli (e.g. stylized images). Here, we present an alternative method that reduces texture bias: pre-training on high-resolution natural images that more closely approximate human visual experience. Our training pipeline needs no data augmentation but solely relies on visual features that occur in everyday scenes. Our method and dataset provide an opportunity to build DCNNs that operate on high-resolution images and may aid in closing the gap between human visual processing and DCNNs.