Enhanced depth estimation by using object placement relation

Abstract

Depth estimation from a scene is an important task in computer vision and 3-d image reconstruction. Normally, human being has an amazing ability to understand a scene quickly by extracting visual information such as object shape, stereo vision cue, size, placement and etc. However, in computer vision, finding 3-d position from an image is still a challenging task, though many researches have been proposed for decades. Many methods have been presented some efficient solutions using image acquisition from both one and several images. Nevertheless, there is no generic solution to recover precise depth from a single image without any prior knowledge. Object placement is one of vision cues usually used to identify 3-d position efficiently, while extraction of such information is not so trivial. Our approach presents an adaptive algorithm defining placement information as a constraint to estimate depth from a single scene image having many arbitrary objects. Our experimental result shows that our algorithm can estimate precise depth from a wide range of image scenes.