Statistics of 3D object locations in images

Abstract

The structural design of buildings and streets in man-made outdoor environments determine the possible locations of other objects like cars or people. Cars drive or park on the streets, and people walk on the sidewalks. Their space of interaction is limited and aligned to these man-made structures. In this report, we describe how 3D locations of objects can be estimated in a single view image. We gather statistics to analyze the locations of cars and people and show how their alignment to man-made structures creates regularities, which can be used to predict their locations. We show which prior knowledge of the analyzed scene is necessary for different ways of statistical prediction of object locations and evaluate the prediction performance on a data set with urban outdoor scenes.