Latent Layout Analysis (LLA) is a novel unsupervised learning technique to discover objects in unseen images using a set of un-annotated training images. LLA defines a generative model that associates latent aspects to local appearances. The dependency between aspects and position is captured by a spatial sensitive aspect model. This dependency distinguishes LLA from Probabilistic Latent Semantic Analysis (PLSA). The latent aspects together with the latent layout constitute a compact scene representation. We demonstrate that the proposed LLA significantly outperforms Probabilistic Latent Semantic Analysis in two tasks: object discovery (detection) and object localization.