Minimum cost aspect classification : a module of a vision algorithm compiler

Abstract: "An important issue in building a model-based vision system is how to extract and organize the relevant knowledge of an object, and systematically turn this knowledge into a working vision system. One approach is to use a vision algorithm compiler, which utilizes stored models of objects, sensors, and processing operations to automatically generate a working vision system. In this paper, we discuss the design of one module of an optimizing vision algorithm compiler which determines the minimum-cost sequence of operations needed to classify an object into an aspect. Given the costs of various feature extraction operations, the module searches over the space of possible classification strategies for the combination of operations that minimizes the expected cost.The optimal strategy is compiled in the form of an aspect classification tree. The classification tree may be expensive to compile, but this cost is incurred off-line, and may result in significant savings at run-time. The performance of the module is illustrated with several examples."