Three dimensional transparent structure segmentation and multiple 3D motion estimation from monocular perspective image sequences

A three dimensional scene can be segmented using different cues, such as boundaries, texture, motion, discontinuities of the optical flow, stereo, models for structure, etc. We investigate segmentation based upon one of these cues, namely three dimensional motion. If the scene contain transparent objects, the two dimensional (local) cues are inconsistent, since neighboring points with similar optical flow can correspond to different objects. We present a method for performing three dimensional motion-based segmentation of (possibly) transparent scenes together with recursive estimation of the motion of each independent rigid object from monocular perspective images. Our algorithm is based on a recently proposed method for rigid motion reconstruction and a validation test which allows us to initialize the scheme and detect outliers during the motion estimation procedure. The scheme is tested on challenging real and synthetic image sequences. Segmentation is performed for the Ullmann's experiment of two transparent cylinders rotating about the same axis in opposite directions.