Parallel simplification and compression of reality captured models

Abstract

Reality capture technologies such as laser scanner and photogrammetry are becoming more democratize day-by-day, as such we are currently undergoing an influx of high-resolution photo-realistic 3D models that contain ever-increasing geometric and texture details and resolution. The massive data file size required to store this information can cause difficulty for both the transfer and manipulation of the captured model. This problem is even more prevalent for mobile or handheld capture devices with limited available internal disk space on the device itself. The three main outputs of a complete reality capture pipeline are dense point cloud, 3d mesh model and texture maps. To reduce the size of each of these outputs, we employ data compression and simplification techniques while striving to retain as much quality as possible. Given the large initial input size and expected long processing time, we explored specifically parallel algorithms in order to fully utilize modern multi-core CPU and GPU to accelerate the computation. This thesis focuses on the parallelized simplification and compression algorithm for large-scale point cloud, 3D mesh, and textures. These input models are generated from running the full reconstruction of a photogrammetry 3D reconstruction pipeline. We have investigated and proposed parallelizable methods to reduce file size for each of the output types while still managed to retain high visual quality like the raw captures.