Authors:

Jeronimo Rodriguez

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Reconstructing semantic 3D building models at Level of Detail 3 (LoD3) remains a complex challenge in the field of urban modeling. Most existing methods focus on LoD2 representations, which lack important façade-level details such as windows and doors details that are essential for applications like urban planning, simulation, and autonomous navigation. While recent advances in deep learning and data fusion have shown progress, further development is needed to effectively combine diverse data types and manage uncertainty in the reconstruction process. This research proposes an approach that uses point cloud data and orthophotos generated from oblique UAV imagery to support the reconstruction of LoD3 building models. The goal is to build a benchmark dataset to test a deep learning model that detects windows directly from raw UAV images and projects them onto existing LoD2 models to enhance them to LoD3. The dataset includes images captured using a DJI M300 drone with a P1 sensor, and the TUM2TWIN dataset is used as a reference to align with current LoD3 reconstruction standards. The proposed pipeline will be tested across several reconstruction strategies using these benchmark datasets. This comparison aims to assess the accuracy and completeness of our method relative to existing approaches. By doing so, we aim to identify where the pipeline performs well, where it may fall short, and how it can contribute to more reliable and interpretable LoD3 modeling workflows. Keywords: Point Cloud; CityGML; Reconstruction; LoD3; UAV

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Jeronimo Rodriguez