Shape Modeling based on Sparse Gaussian Process Implicit Surfaces

Conference poster, 2018

Reconstructing, modeling, and accounting for uncertainty in three-dimensional shapes is important in a large
number of areas, such as biometrics, biomedical imaging, data mining, robotics. However, it is challenging to build accurate models of novel objects based on real sensory data as the measurements are often incomplete and noisy. Besides, imperfect sensory data requires explicit uncertainty modeling that can enable action planning with maximum information gain and efficient use of data. We present a probabilistic approach for learning object models based on visual and tactile data. We study Gaussian Process Implicit Surface (GPIS) representation, a non-parametric probabilistic reconstruction of object surfaces from 3D data points which provides a principled approach to encode uncertainty in the data, and investigate different configurations for GPIS. We interpret an object surface as the level-set of an underlying sparse GP. Experiments are performed on synthetic and real data sets obtained from physical interaction with objects. We evaluate results by assessing how close the reconstructed surfaces are to the ground truth, and how well objects from different categories are clustered based on the obtained representation. Results show that sparse GPs enable a reliable approximation to the full GP solution and the proposed method yields adequate surface representations to distinguish objects.