Speaker: Philipp Erler
Abstract
Point clouds, while easy to acquire, are often noisy, incomplete, and lack connectivity. They cannot represent closed surfaces efficiently, and processing for geometrical applications like 3D printing is difficult. Therefore, we want to convert them to surface representations, most importantly, triangle meshes. This dissertation addresses the challenge of reconstructing accurate and robust 3D objects from imperfect point clouds.
Traditional methods struggle with point cloud defects, motivating the exploration of data-driven approaches. However, deep learning on point clouds is difficult due to their unordered and unstructured nature. Further, subsampling is necessary to fit them into the fixed-size input of a neural network, which introduces randomness. In this thesis, we explore solutions for these challenges.
We present three main contributions: Points2Surf, which combines local and global priors to generalize surface reconstruction across diverse object classes; PPSurf, which leverages point convolutions and attention to further improve reconstruction quality; and LidarScout, which enables real-time, out-of-core rendering of massive aerial LIDAR scans by focusing on efficient, local heightmap estimation.
Our results demonstrate that balancing local detail with global context is key to achieving high-quality, generalizable reconstructions. We show that model design, data augmentation, and efficient representations are crucial for handling noise and missing data. The findings offer practical solutions for reconstructions from single objects to entire landscapes.