The architecture of a distributed vision system is presented, with particular attention directed to the bottom-up indexing mechanism performed by a hierarchically organized network of information processing (IP) modules. Each IP module adaptively transforms input data passed by lower-level modules into more complete observations and performs a transformation that is modeled as a regularization process. This scheme is applied to the problem of recognizing objects whose 3-D shape can be described as a set of planar surfaces. Edge detection, straight-line extraction, grouping, and matching are the P modules considered. In particular, the regularization process consists of either a voting scheme or a Markov random field labeling process, depending on the level. At the higher level, a degree of belief is given about the presence of objects contained in the scene and considered in the model database. Results demonstrate both the validity of the processes applied separately at each level and the global consistency of the method.