Modelling Data Complexity for Model-based Vision

This paper discusses an issue of wide-ranging importance for computer vision the systematic consideration of data complexity in assessment of computer vision systems. We investigate 3D object recognition from 2D features as a typical problem of computer vision. We identify 3 factors which contribute to the complexity of image data: feature truncation, noise and clutter. We propose a modelling scheme for these factors, which allows us both to measure and to simulate each factor. Using the scheme, a systematic comparison is made between two existing strategies for model-matching as a function of clutter and truncation factors. Post-model-matching object discrimination is then examined as a function of the noise factor. These two examples serve to illustrate the data complexity model, and demonstrate its use for formal assessment of modelbased algorithms.