In the setting of computer vision, algorithmic searches often aim to identify an object of interest inside large sets of images or videos. Towards reducing the often astronomical complexity of this search, one can use pruning to filter out objects that are sufficiently distinct from the object of interest, thus resulting in a pruning gain of an overall reduced search space. Motivated by practical computer vision based scenarios such as time-constrained human identification in biometric-based video surveillance systems, we analyze the stochastic behavior of time-restricted search pruning, over large and unstructured data sets which are furthermore random and varying, and where in addition, pruning itself is not fully reliable but is instead prone to errors. In this stochastic setting we apply the information theoretic method of types as well as information divergence techniques to explore the natural tradeoff that appears between pruning gain and reliability, and proceed to study the typical and atypical gainreliability behavior, giving insight on how often pruning might fail to substantially reduce the search space. The result, as is, applies to a plethora of computer vision based applications where efficiency and reliability are intertwined bottlenecks in the overall system performance, and the simplicity of the obtained expressions allows for rigorous and insightful assessment of the pruning gain-reliability behavior in such applications, as well as for intuition into designing general object recognition systems.
展开▼
