Abstract

This paper proposes to infer accurately a 3D shape of an object captured by a depth camera from multiple view points. The Generalised Relaxed Radon Transform (GR2T) [1] is used here to merge all depth images in a robust kernel density estimate that models the surface of an object in the 3D space. The kernel is tailored to capture the uncertainty associated with each pixel in the depth images. The resulting cost function is suitable for stochastic exploration with gradient ascent algorithms when the noise of the observations is modelled with a differentiable distribution. When merging several depth images captured from several view points, extrinsic camera parameters need to be known accurately, and we extend GR2T to also estimate these nuisance parameters. We illustrate qualitatively the performance of our modelling and we assess quantitatively the accuracy of our 3D shape reconstructions computed from depth images captured with a Kinect camera.