CompleteRoof: Point cloud completion with deep learning (Student Research Assistant 8h/w)

16.07.2024, Studentische Hilfskräfte, Praktikantenstellen, Studienarbeiten

We are looking for a student research assistant for establishing new benchmark and methods in point cloud completion tasks.

The full description: https://syncandshare.lrz.de/getlink/fiYRWWdnER7yS9nz5sfugS/CompleteRoof_hiwi.pdf

Description
Acquiring high coverage and high-quality point clouds has
been a long-standing challenge in photogrammetry, remote sensing, and
computer vision. In real-world scenarios, point cloud capturing sensors are
prone to measurement noise and occlusions, frequently rendering the point
clouds corrupted and incomplete. This phenomenon is particularly apparent
when acquiring aerial-based point clouds of buildings, where roofs are often
tree-occluded and noisy. Complete point clouds of buildings are pivotal to
multiple downstream tasks, such as semantic 3D building reconstruction for
solar potential analysis [1].
While there exist methods of completing 3D point clouds [2], the comple-
tion of 3D roofs remains in its infancy. They perform well on symmetrical
shapes but are limited to more complex and highly occluded roof types.
Notably, the current methods are primarily tested at the regional and city
scales, neglecting the variability of roof types at a country-specific level.

Objectives
In this project, you will be supervised by Olaf Wysocki (Pho-
togrammetry & Remote Sensing) and work closely with Benedikt Schwab
(Geoinformatics) and Dr Yan Xia (Computer Vision). This position is for
8 hours/week, starting in August/September until the end of the semester,
March/April, with the possibility of extension; The joint publication at a
top venue is foreseen. Your tasks will include:
• Acquiring open data point clouds, creating benchmark data;
• Adapting existing point cloud completion methods for the roof com-
pletion;
• Establishing baselines using the state-of-the-art methods for point
cloud completion.

Requirements Sound programming skills in Python are required. Familiarity with libraries such as OpenCV,
Open3D, and PyTorch is highly advantageous. Understanding machine learning, computer vision, and pho-
togrammetry fundamentals is required. Student status, CV, and the current transcript of records are a must
(send to olaf.wysocki@tum.de).

References
[1] Wysocki, O., Xia, Y., Wysocki M., Grilli, E., Hoegner, L., Cremers D., and Stilla, U.
Scan2LoD3: Reconstructing semantic 3D building models at LoD3 using ray casting and Bayesian networks,
CVPRW 2023, https://shorturl.at/qzSX0
[2] Lo, KSH., Peters, J., Spellman, E., RoofDiffusion: Constructing Roofs from Severely Corrupted Point Data via Diffusion, ECCV 2024, https://arxiv.org/pdf/2404.09290

Kontakt: olaf.wysocki@tum.de

More Information

https://syncandshare.lrz.de/getlink/fiYRWWdnER7yS9nz5sfugS/CompleteRoof_hiwi.pdf