|Ph.D Student||Berengolts Alexander|
|Subject||Quantitative Approach for Characterizing the Difficulties|
of Visual Tasks
|Department||Department of Computer Science||Supervisor||Professor Michael Lindenbaum|
In recent years the necessity of performance evaluation of perceptual grouping algorithms became evident. In this thesis we provide, for the first time, a quantitative analytic performance prediction for grouping algorithms.
We first analyze the performance of a grouping process based on the connected components algorithm. We derive the expected number of addition errors and the group fragmentation rate. We show that these performance figures depend on a few inherent and intuitive parameters. We further show that it is possible to control the grouping process so that the performance may be chosen within the bounds of a given tradeoff. The analytic results are supported by implementing the algorithm and testing it on synthetic and real images.
Then we propose here a modified saliency estimation mechanism, which is based on probabilistically specified grouping cues and on curve length distributions. The proposed approach lends itself to different types of generalizations, and in particular to saliencies based on different cues, in a systematic rigorous way. To demonstrate that, we created a saliency process based on gray level similarity.
We show however that only a limited class of saliency function may be rigorously optimized by a local process. Following this result we focus on probabilistic analysis of expected length saliencies. Using ergodicity and asymptotic analysis, we derive the saliency distribution associated with the main curves and with the rest of the image. We then extend this analysis to finite-length curves. Based on the derived distributions we show how to set a threshold on the saliency for optimally deciding between figure and background, how to choose cues which are usable for saliency, and how to estimate bounds on the saliency performance.