Skip to main navigation menu Skip to main content Skip to site footer

Handling Colors in Image Classification


The presence of bias in a dataset has been a long standing bottleneck in the task of image classification. While supervised methods have shown to overcome these biases, self-supervised methods have managed to overcome benchmarks set by supervised learning methods. This paper shows that self-supervised methods can maintain their ability to outperform supervised methods even when introduced to color bias. Two experimentation pipelines are presented. One focuses on the capability of a model to handle artificially induced color bias and the other gauges the ability of a model to incorporate naturally occurring color differences present in vision datasets.