Player detection is an important task in sport video analysis. Once players are detected accurately, it can be used for player tracking, player activity/performance analysis as well as team activity recognition. Recently, convolutional Neural Networks (CNN) became the state-of-the-art in computer vision for object recognition. CNN based methods usually use gray or RGB images as an input. It is also possible to use other image representation techniques such as shape information image and polar transformed shape information image for player detection. In this paper, we evaluate various image representation techniques for player detection using CNN. In our evaluation, first the candidate image regions for players are determined using a sliding window technique. Then these regions are input to CNN for player detection. We examine four different types of image representations as an input to CNN: RGB, gray, shape information and polar transformed shape information image. Evaluation is conducted on a field hockey dataset. Results show that CNN based player detection is effective and different image representations yield different performances.