Colonoscopic 3D reconstruction by tubular non-rigid structure-from-motion

The visual examination of colonoscopic images fails to extract precise geometric information the colonic surface. Reconstructing 3D surface colon from image sequences may thus add valuable clinical information. We address this problem extracting spatio-temporal structure images. Using just intrinsically calibrated monocular stream, we develop a technique compute depth certain feature points that have been tracked across Our method uses prior knowledge an approximate geometry colon, called (TTP). It works by fitting deformable cylindrical model reconstructed independently non-rigid structure-from-motion (NRSfM), compromising between data term and novel tubular smoothing prior. represents first ever exploit very weak topological improve NRSfM. As such, it lies in-between standard NRSfM, which does not use beyond mere plane, shape-from-template (SfT), strong as full object model. validate our on both synthetic structures real data. improves results obtained existing NRSfM methods 71.74% average succeeds in obtaining reconstruction sequence defeating methods. Colonoscopic is difficult problem, yet unresolved computer vision. proposed dedicated experiments show motion might be right cue colonoscopy.