Recently, sparse coding based on unsupervised learning has been widely used for image classification. The sparse representation is assumed to be linearly separable, and therefore a simple classifier, like softmax regression, is suitable to perform the classification process. To investigate that, this paper presents a novel approach for semantic place recognition (SPR) based on Restricted Boltzmann Machines (RBMs) and a direct use of tiny images. These methods are able to produce an efficient local sparse representation of the initial data in the feature space. However, data whitening or at least local normalization is a prerequisite for these approaches. In this article, we empirically show that data whitening forces RBMs to extract smaller structures while data normalization forces them to learn larger structures that cover large spatial frequencies. We further show that the later ones are more promising to achieve the state-of-the-art performance for a SPR task.