Colour is important, but how important? This study addresses the question by testing a deep learning approach, ResNet-50, on the task of object classification based on using full-colour, dichromatic, and grayscale images as inputs and comparing the recognition performance as the amount of colour information is reduced. The results show that colour is useful, but far from crucial for object classification. The error rate increases by only 12% for the grayscale case over the full-colour case. A examination of some of the cases in which the full-colour classifier succeeds, but the grayscale classifier fails, reveals the interesting trend that while the in some cases the colour features of an object are crucial, colour may be perhaps even more important for understanding occlusion ordering and figure-ground separation.