3D Geometric Shape Modeling by ‘3D Contour Cloud’ Reconstruction from Stereo Videos

We present a shape representation for objects based on 3D contour fragments that build a ‘3D Contour Cloud’. Our approach for the automatic reconstruction of such ‘3D Contour Clouds’ from calibrated stereo image sequences includes a novel idea on the reconstruction of 3D contour fragments from stereo frame pairs and a robust Structure and Motion analysis. Moreover, we propose an extension of 2D shape context to 3D – 3D shape context – which we use for outlier analysis. We show results on a standard multi-view stereo dataset as well as on our own stereo dataset. We achieve qualitatively convincing ‘3D Contour Clouds’ for various objects including the automatic generation of 3D bounding boxes and object-centered coordinate systems.