Consistent labelling of image features using an assumption-based truth maintenance system

摘要

Labelling sets of 2D image features as model features is a constraint satisfaction problem that occurs in model-based vision. The labelling must be consistent with constraints that describe how image features originating from the modelled object would appear in the image. This paper discusses how an assumption-based truth maintenance system (ATMS) can be used to solve such a constraint-satisfaction problem. The ATMS is used to limit the number of constraints applied, and to represent the multiple sets of consistent labels that are possible. The effectiveness of the ATMS in limiting the constraints is analysed.