Abstract
Point clouds are a set of data points in space that represent an object in a 3D coordinate system. The use of point cloud data for sub-sea surveys for maintenance of under-water assets have received increased attention over the last few years. However, point cloud data acquired in a sub-sea environment is full of holes and noises due to the adverse visibility conditions of the marine environment. Point cloud completion, the task of generating complete point cloud data from a partial input, can help fill the gaps so that the information important for asset maintenance can be communicated. Unfortunately, even the best state-of-the-art-completion algorithms have great limitations when it comes to performing completion on real-world datasets. This dissertation presents a literature survey of modern completion algorithms in light of their usefulness in completing real-world under-water point clouds and performs an experiment on a leading completion algorithm, PoinTr, with both synthetic and real-world datasets. Then, it draws some insights from the literature survey and the experiment and presents potential human-in-the-loop approaches in point cloud completion for the survey and maintenance of under-water assets.