Abstract
We present a learning-based method for interpolating and manipulating 3Dshapes represented as point clouds, that is explicitly designed to preserveintrinsic shape properties. Our approach is based on constructing a dualencoding space that enables shape synthesis and, at the same time, provideslinks to the intrinsic shape information, which is typically not available onpoint cloud data. Our method works in a single pass and avoids expensiveoptimization, employed by existing techniques. Furthermore, the strongregularization provided by our dual latent space approach also helps to improveshape recovery in challenging settings from noisy point clouds across differentdatasets. Extensive experiments show that our method results in more realisticand smoother interpolations compared to baselines.