Building LOD Representation for 3D Urban Scenes

TL;DR

This paper introduces the LOD-Tree, a structure-aware, coarse-to-fine representation for generating semantic LODs of 3D urban scenes from dense geometry. A novel Inside/Outside View (IO-View) analysis identifies principal versus secondary planar primitives and forms level sets that drive a semantic BSP-based partition, producing a set of anchor and interpolation models via a diff-value guided traversal. The approach is validated on 21 real-world datasets, showing improved emergence order of critical structures, robustness to noise, and superior semantic-geometry trade-offs compared with BSP, Lowpoly, NeuralLOD, QEM, and Robust-lowpoly, with a user-friendly interactive selection of LODs. The work enables efficient, semantically meaningful LOD extraction suitable for urban modeling tasks and downstream applications, while acknowledging limitations in plane detection noise and editing capabilities and pointing to future integration with multiple plane-detection configurations and image-based cues.

Abstract

The advances in 3D reconstruction technology, such as photogrammetry and LiDAR scanning, have made it easier to reconstruct accurate and detailed 3D models for urban scenes. Nevertheless, these reconstructed models often contain a large number of geometry primitives, making interactive manipulation and rendering challenging, especially on resource-constrained devices like virtual reality platforms. Therefore, the generation of appropriate levels-of-detail (LOD) representations for these models is crucial. Additionally, automatically reconstructed 3D models tend to suffer from noise and lack semantic information. Dealing with these issues and creating LOD representations that are robust against noise while capturing the semantic meaning present significant challenges. In this paper, we propose a novel algorithm to address these challenges. We begin by analysing the properties of planar primitives detected from the input and group these primitives into multiple level sets by forming meaningful 3D structures. These level sets form the nodes of our innovative LOD-Tree. By selecting nodes at appropriate depths within the LOD-Tree, different LOD representations can be generated. Experimental results on real and complex urban scenes demonstrate the merits of our approach in generating clean, accurate, and semantically meaningful LOD representations.

Figures (24)

• Figure 1: Overview: Our method takes a dense triangle mesh or point cloud as input (a) and performs an IO-View analysis to analyze the properties of planar primitives. This analysis automatically groups these planar primitives into multiple level sets, as shown in different colors in (b). We then sequentially partition the space using these level sets to generate a LOD-Tree (c). Each selected combination of tree nodes corresponds to an output model. We showcase six representative models (d) derived from the LOD-Tree.
• Figure 2: The effect of the $\alpha$-value on the $\alpha$-shape. As the $\alpha$-value goes higher, the hole on the planar primitives can be filled. The Top row shows the boundary changes of a model composed of multiple planes, while the bottom row shows the boundary changes of a single plane.
• Figure 3: We first detect multiple $\alpha$-shapes. Each colored line corresponds to an $\alpha$-shape; as shown in the upper left. These $\alpha$-shapes split the space into polyhedra, and two adjacent polyhedra share a common face; as shown in the bottom left. By removing faces that are not covered by an $\alpha$-shape, we can divide the entire space into four categories, as shown in the middle. Finally, we extract the primitives shared by "core interior" and "core exterior" as the principal primitives $S_{0}$; see the colored lines on the right figure. The remaining planar primitives are labelled as secondary primitives, which define addon (3 pinkish lines) and cutout (3 blueish lines) structures.
• Figure 4: Given three level sets $\{S_0, S_1, S_2\}$ (blue, green and red), we gradually get the LOD-Tree divided by these three sets and generate multiple in/out labels (Label 0, Label 1, Label 2) for leaf nodes to facilitate the construction of LOD-Tree. The label $k$ of the leaf nodes is determined by a majority vote based on level sets $\bigcup_{i=0}^k S_i$. The diff-value of each node depends on the similarity of the space associated with this node and all its leaf nodes by comparing their labels and sizes. That is, node $A$ is labelled as $in$, while leaf nodes of $A$ are labelled as $A'$ (circled in black), then we compute the difference of $A$ and $A'$ as the diff-value of $A$. A node with a higher diff-value is considered less similar to the finest model and, therefore, more valuable when splitting the space. In particular, if a cluster contains fewer than 10 primitives per structure, we will perform a merging operation to merge BSP nodes that are split by the cluster's primitives into a single LOD-node (e.g., windows, chimneys). That is, when we reach node $C$, if we want to cut it next time, node $C$ will be expanded into five leaf nodes at one time. For simplicity, multiple windows are represented by a single window in the diagram, and the same goes for chimneys.
• Figure 5: Illustration of the evolution of priority queue $Q$ and LOD model extraction. Interpolation models ($M^I$) are extracted when the appearance changes significantly. Anchor models ($M^A$) are extracted when each level set $S_i$ completes space partitioning. Both model types are stored in the candidate LOD model set $M$.