The computer vision community has greatly benefited from transferring techniques originally developed in the document processing domain to the visual domain by means of discretizing the features space into visual words. This paper reinvestigates the necessity of this artificially discretization of the continuous space of visual features and consequently proposes an alternative formulation of the popular topic models that is based on kernel density estimates. Results indicate the benefits of our model in terms of decreased perplexity as well as improved performance on object discovery tasks.