The retrieval facilities of most Peer-to-Peer (P2P) systems are limited to queries based on a unique identifier or a small set of keywords. The techniques used for this purpose are hardly applicable for content-based image retrieval (CBIR) in a P2P network. Furthermore, we will argue that the curse of dimensionality and the high communication overhead prevent the adaptation of multidimensional search trees or fast sequential scan techniques for P2P CBIR. In the present paper we will propose two compact data representations which can be distributed in a P2P network and used as the basis for a source selection. This allows to communicate only with a small fraction of all peers during query processing without deteriorating the result quality significantly. We will also present experimental results confirming our approach.