Automatic generation of image analysis programs

Abstract

In this paper, we introduce a system that generates computer vision programs for a given task, which is specified by regions of interest in a collection of example images. The system relies on a database of operators, which are combined by an automated planning approach in order to create executable programs. We present an early proof-of-concept implementation that relies on a limited database to solve simple tasks, such as finding players in a soccer video or cups on a table. Our experimental evaluation shows that the basic approach is working on relative simple scenarios. Future work will focus on integrating more complex problem descriptions, which require more sophisticated planning strategies in order to compensate for rapidly increasing search spaces.