Girgit: A Dynamically Adaptive Vision System for Scene Understanding

Abstract

Modern vision systems must run in continually changing contexts. For example, a system to detect vandalism in train stations must function during the day and at night. The vision components for acquisition and detection used during daytime may not be the same as those used at night. The system must adapt to a context by replacing running components such as image acquisition from color to infra-red. This adaptation must be dynamic with detection of context, decision on change in system configuration, followed by the seamless execution of the new configuration. All this must occur while minimizing the impact of dynamic change on validity of detection and loss in performance. We present Girgit, a context-aware vision system for scene understanding, that dynamically orchestrates a set of components. A component encapsulates a vision-related algorithm such as from the OpenCV library. Girgit inherently provides loading/caching of multiple component instances, system reconfiguration, management of incoming events to suggest actions such as component re-configuration and replacement of components in pipelines. Given the surplus architectural layer for dynamic adaptation one may ask, does Girgit degrade scene understanding performance? We perform several empirical evaluations on Girgit using metrics such as frame-rate and adaptation time to answer this question. For instance, the average adaptation time between change in configurations is less than 2 μs with caching, while 8 ms without caching. This in-turn has negligible effect on scene understanding performance with respect to static C++ implementations for most practical purposes.