Put things in correct location: Describing the scene with contextual cues

摘要

In this paper, we reviewed the development of scene understanding, specifically focusing on indoor scene understanding and context used in scene understanding. Different from image segmentation, object detection, 3D reconstruction, etc., which only implements a single task, scene understanding is a technology aiming at implementing a holistic scene parsing, that is, acquiring the spatial extent, location and semantic of every object in a scene. Nevertheless, scene understanding doesn't equal to direct accumulation of individual vision tasks. Along with the development of scene understanding, it has formed its own framework. In general, indoor scene understanding is more challenging than outdoor scene understanding, but with the emergence of RGB-D sensors, more and more researchers have focused on indoor scene understanding. Fusing context information to constrain relationship between adjacent pixels and improve accuracy of algorithm is the common sense in the field of scene understanding. There are three kinds of context, namely, pixel based context, region based context and object based context. In this paper, We reviewed these different context respectively.