Abstract
The AI task of polyhedral scene analysis from line-drawings was undertaken using a network approach. A harmony theory network was constructed, where the two obvious environmental levels of the input (lines and junctions) served as the primitives for the network's two layers. The implementation was found to produce (one of) the correct solution(s) for the given object, subject only to probabilistic errors. It also exhibited the following interesting properties.
1 Additivity under decompositions, where classification was carried out independently for parts of the scene between which no constraints flowed (separate objects or occluded and occluding ones)
2 Completion of partly labelled or of unlabelled images
3 Robustness, since in the case of non-labellable objects the end state invariably involved the minimum number of violations of the precompiled knowledge