Building models from multiple point sets with kernel density estimation
McDonagh, Steven George
One of the fundamental problems in computer vision is point set registration. Point set registration finds use in many important applications and in particular can be considered one of the crucial stages involved in the reconstruction of models of physical objects and environments from depth sensor data. The problem of globally aligning multiple point sets, representing spatial shape measurements from varying sensor viewpoints, into a common frame of reference is a complex task that is imperative due to the large number of critical functions that accurate and reliable model reconstructions contribute to. In this thesis we focus on improving the quality and feasibility of model and environment reconstruction through the enhancement of multi-view point set registration techniques. The thesis makes the following contributions: First, we demonstrate that employing kernel density estimation to reason about the unknown generating surfaces that range sensors measure allows us to express measurement variability, uncertainty and also to separate the problems of model design and viewpoint alignment optimisation. Our surface estimates define novel view alignment objective functions that inform the registration process. Our surfaces can be estimated from point clouds in a datadriven fashion. Through experiments on a variety of datasets we demonstrate that we have developed a novel and effective solution to the simultaneous multi-view registration problem. We then focus on constructing a distributed computation framework capable of solving generic high-throughput computational problems. We present a novel task-farming model that we call Semi-Synchronised Task Farming (SSTF), capable of modelling and subsequently solving computationally distributable problems that benefit from both independent and dependent distributed components and a level of communication between process elements. We demonstrate that this framework is a novel schema for parallel computer vision algorithms and evaluate the performance to establish computational gains over serial implementations. We couple this framework with an accurate computation-time prediction model to contribute a novel structure appropriate for addressing expensive real-world algorithms with substantial parallel performance and predictable time savings. Finally, we focus on a timely instance of the multi-view registration problem: modern range sensors provide large numbers of viewpoint samples that result in an abundance of depth data information. The ability to utilise this abundance of depth data in a feasible and principled fashion is of importance to many emerging application areas making use of spatial information. We develop novel methodology for the registration of depth measurements acquired from many viewpoints capturing physical object surfaces. By defining registration and alignment quality metrics based on our density estimation framework we construct an optimisation methodology that implicitly considers all viewpoints simultaneously. We use a non-parametric data-driven approach to consider varying object complexity and guide large view-set spatial transform optimisations. By aligning large numbers of partial, arbitrary-pose views we evaluate this strategy quantitatively on large view-set range sensor data where we find that we can improve registration accuracy over existing methods and contribute increased registration robustness to the magnitude of coarse seed alignment. This allows large-scale registration on problem instances exhibiting varying object complexity with the added advantage of massive parallel efficiency.