Reasoning about objects in a visual scene can be substantially improved if we have a representation that includes both information about the object properties and information about the relations that hold between them. For this, we have built a semantic-relational network that makes use of textual commonsense knowledge of both sorts and allows for inference within a Bayesian framework (Roehrbein, Eggert, & Koerner, 2007). In order to enrich this graphical representation, we target at adding complementary information gained, for example, from sets of labeled images. We are especially interested in relational knowledge, a demand that rules out most existing image databases since they usually contain only one labeled object per image. LabelMe proved as a promising alternative and we started our efforts by analyzing the statistical dependencies between all objects in fully-labeled images of that database. In the contribution here, we report these results and discuss how the gained information can be used to build contexts as a first important step in reaching the final goal of learning a relational knowledge representation in an autonomous way.