In this paper, we propose a novel approach for the image-based classification of ancient coins by recognizing motifs minted on their reverse sides. Reverse motif recognition is achieved with dense sampling-based bag-of-visual-words model. As the bag-of-visual-words model lacks spatial information of visual words, we evaluate three types of tiling schemes to incorporate such information. Due to the specific geometric structures of the objects in the reverse motifs, spatial information is also important for a better motif recognition rate. Furthermore, a coin can be imaged under various rotations. We also investigate the type of spatial scheme more robust to such rotations. Other parameters such as the size of visual vocabulary, resolution of the dense sampling grid, number of features collected per image to construct the visual vocabulary and the number of tilings in each tiling scheme are also investigated.