Abstract: The work presented is part of a larger effort to build a general object recognition system. Objects as well as human faces are represented by graphs labeled with Gabor filter responses. We describe an optimal method to reconstruct images from such graphs. Two examples of how this can be used to analyze the object representation or to compensate for its deficiencies are presented. Since the reconstruction method is formulated generally for an arbitray set of linear filters, it can also be applied to data produced by other systems, artificial or biological.