Open VL: Abstracting Vision Tasks Using a Segment-Based Language Model

摘要

Computer vision is a complex field which can be challenging for those outside the community to apply in the real world. In this paper we show one method to provide access to sophisticated computer vision methods to general developers, hobbyists or researchers outside the field. Our contribution is an abstraction utilising fundamental vision operations based on a single unit, the segment, can be used to describe images, local image conditions and between-image conditions. We illustrate how a descriptive language model can be built on the segment to provide an intuitive mental model of computer vision to mainstream developers. We demonstrate how we can map a description of the task composed of the segment-based language into the space of algorithms, to choose an appropriate method to solve the problem. We use the problems of segmentation, correspondence and image registration to show how end-to-end problems may be constructed using our novel metaphor.