Abstract: Most deep-learning models for vision tasks rely on RGB images as their primary input layer, assuming the model inherently discovers an optimal representation. In this work, we challenge this ...