MULTI-VIEW IMAGE INTERPRETATION FOR STREET-SIDE SCENES

摘要

In this paper we examine the concept of redundancy and how it can improve the scene interpretation. In our work, we focus on redundant sets of street-side images. Semantic segmentation is performed on each image. Results of the segmentation are compared in overlapping images and matched. We use two principally different datasets to validate our results. The Industrial System dataset is taken from a moving car by well-designed, calibrated, automated cameras, with the geometry and pattern of the images accurately defined. Our second dataset (Tummelplatz-Graz) was taken by a hand-held camera in an urban environment, following the "crowd-sourcing" paradigm. Each database provides its typical level of redundancy and different approaches are needed for image matching. The annotated Tummelplatz-Graz database will be also released for public to make further references and comparison easier.