Scene parsing using graph matching on street-view data

• An effective scene parsing framework via graph matching guidance on street-level data is proposed.
• Graph matching is introduced to partially match image components taking into account the regional similarity of scenes.
• The proposed algorithm can be applied to small training and testing sets, and achieves competitive parsing performance.

Scene parsing, using both images and range data, is one of the key problems in computer vision and robotics. In this paper, a street scene parsing scheme that takes advantages of images from perspective cameras and range data from LiDAR is presented. First, pre-processing on the image set is performed and the corresponding point cloud is segmented according to semantics and transformed into an image pose. A graph matching approach is introduced into our parsing framework, in order to identify similar sub-regions from training and test images in terms of both local appearance and spatial structure. By using the sub-graphs inherited from training images, as well as the cues obtained from point clouds, this approach can effectively interpret the street scene via a guided MRF inference. Experimental results show a promising performance of our approach.