The model-based vision approach also has an early knowledge-based examplar, ACRONYM [10], which used symbolic reasoning to aid static scene interpretation. WALKER [29] was an early dynamic model-driven interpretation system that could identify examples of moving people in image sequences. In model-based vision, the stored knowledge is concerned with the expected objects, often specifying part-whole relationships and constraints among the subparts, but also relationships over time. The visual processing is driven by hypotheses, primarily top-down. For example, the ACRONYM system used stored models in the form of slot and filler frames which formed the nodes of the ``object graph''. Generalised cylinders were used as primitives in this hierarchical structure which represented objects from coarse to fine detail. Algebraic constraints could also be specified to build up the hierarchical ``restriction graph''. To drive the processing, ACRONYM constructed a ``prediction graph'' using these models and some reasoning. Then low-level edge and ribbon-like structures were constructed under the direction of the predictor module to form the ``observation graph''. Finally, the ``interpretation graph'' matched the observed features and relationships to the models using more reasoning to eliminate inconsistencies. Again, more recently, model-based vision systems have been refined using probabilistic techniques, for example [5].
Model-based vision techniques have also been refined by Koller and Nagel [39] using fully parameterised object models which can deliver detailed descriptions of tracked objects. Another important technique is to use 2D iconic representations from different views of the 3D model to simplify the matching. For example, Sullivan and colleagues [65,70] have developed model-based tracking in traffic scenes for performance under real-time constraints. There is ongoing debate about the roles of iconic and 3D representations in the many different tasks performed by computer vision systems. Another notable development in model-based vision is the use of deformable objects which have to be described using statistical rather than geometric relationships [17,64]. A major advantage of such representations is that they can be learnt from examples, for example in work by Baumberg and Hogg [3]. The use of iconic representations and statistical relationships, which can easily be acquired from images, is generally accepted to be biologically plausible. However, there are many open questions about the effectiveness of more formal analysis and the modelling of high-level invariance for computer vision tasks.