Estimation of relative depth in the scene using SFF-inspired focus cue

Abstract

Three dimensional (3-D) vision techniques in the field of Computer Vision aims mainly at reconstructing a scene (surface) to find its three dimensional geometrical information from 2D image(s) and sometimes with the aid of some special devices. Passive 3-D vision techniques such as computational stereo vision method do surface reconstruction from disparities arising in images of the same scene taken from multiple views. Shape from Focus (SFF) is a method which recovers the 3D geometry of the scene based on a sequence of images taken from different focus distances between the camera and the object. Generally SFF techniques require parallel projection of the scene on to the image plane so that the corresponding pixels in the set of images taken are easily identified. This can be achieved by using a lens which does parallel projection such as a telecentric lens or a microscopic lens. Moreover the SFF method is widely applied for extremely small objects due to the limited range of magnification that can be maintained. This again is another manifestation of the fact depth of objects produce perspective shift (generally called as structure-dependent pixel motion) in the image plane. All these facts are applicable for situations which utilizes SFF for complete reconstruction of the scene. Applications involving shape information extracted from focus as a secondary cue need not require a complete dense reconstructed information from SFF. Such applications might allow usage of wide angle lenses where the projection is basically a perspective projection of the scene on to the image plane. The paper utilizes a wide angle lens for SFF based scene reconstruction consisting of a macroscopic object. The reconstruction made using the SFF inspired focus cue may be used as a deciding factor for the baseline in a single moving camera based stereo vision system. The method is validated for objects with different texture and heights.