Lightweight 3D Modeling of Urban Buildings From Range Data

Lightweight 3D Modeling of Urban Buildings From Range Data Weihong Li* George Wolberg* Siavash Zokai+ * The City University of New York + Brainstorm Technology LLC

Outline • Introduction • Overview • Methodologies – 3D point cloud preprocessing – Lightweight 3D reconstruction – Model generation

• Experimental Results • Performance Evaluation • Conclusion and Future Work The City University of New York

Introduction • Background – Massive point clouds by laser range scanners present challenges to efficient modeling, visualization and storage. – State-of-the-art techniques are not scalable on large datasets – Lack of techniques for web-based applications

• Related Work – Procedural Modeling of Urban Buildings – Reverse Engineering – Image-based modeling

• Contribution – A Framework for Lightweight 3D Reconstruction – Cost-Effective Geometry Compression – Produce Wide Spectrum of Resolutions

The City University of New York

Overview Data acquisition

Slab partitioning

Slab projection Volumetric slabs

Point cloud

2D Slice

Keyslice detection / Boundary vectorization

Tapering detection

Rendering

Output model

Tapered model

Extrusion

Extruded model

The City University of New York

Keyslices contours

Flow Diagram Pre-processing 3D Data Generation & Acquisition

Major Plane Detection

2D Slice Extraction & Enhanceme nt

Dataset Segmentatio n

Boundary Vectorization

Tapering Detection

Lightweight 3D Reconstruction Window/Door Detection

Keyslice Detection

Y

More Segments?

N Post-processing Model Generation

Model Merging

Window/Door Installation

3D Data Generation and Acquisition • Synthetic Dataset Generation

• Real Dataset Acquisition

3D Point Cloud Preprocessing • Major Planes Detection – Moving least square based local methods – Point cloud data rectification

• 2D Slice Extraction and Enhancement – Slice Extraction along the normal of major planes – Hole filling for missing data

• Dataset Segmentation – Separators computation – 2D slices segmentation

The City University of New York

Major Plane Detection • Approaches on Plane Detection of 3D Points – Plane Fitting – Hough Space Transform

• Normal Estimation on 3D Points – Local plane fitting based on MLS optimization – Implemented with kd tree – With complexity of O(nlogn)

• Point Cloud Data Rectification – Apply matrix M to each 3D point P(x,y,z) , P’ = MP – M is updated for each major normal

The City University of New York

2D Slice Extraction 

• Slice Extraction – Along the normal of major planes

 

Y

[ x 2 D , y 2 D ]T    [ xi3D  X MIN , zi3D  Z MIN ]T

Z

 X

2D Slice Extraction and Enhancement • Slice Extraction – Along the normal of major planes

• Hole Filling for Missing Data Recovering – Symmetry detection – Lightweight 2D computation

L  arg min  d x , y ( P' , I ) x, y

The City University of New York

Dataset Segmentation • Approaches on Data Segmentation – Edge based methods – Region grow based methods

• Separators Based Segmentation – Salient features

The City University of New York

Dataset Segmentation • Separators Based Segmentation – 2D and 3D segmentation

Lightweight 3D Reconstruction • Window and Door Detection – Window and door computation – Mask image generation

• Keyslice Detection – Distance measurement based – Curvature based

• Boundary Vectorization – 2D adaptive BPA – Contour simplification

• Tapering Detection – Linear tapering to point/line/offset – Tapering structure verification

The City University of New York

Window and Door Detection • Window and Door Computation – For images (Lee et al. CVPR04) and for PCDs (Pu et al. ISPRS07) – Compute the depth and locations of windows

The City University of New York

Keyslice Detection • Extrusion Computation – Identify keyslices based on similarity measure – Hausdorff distance N

d H ( I , I r )   d min ( Pi , I r ) I is keyslice ifd H ( I , I r )   d i 0

Boundary Vectorization • Adaptive Ball Pivoting Algorithm (ABPA) – Initial BPA contour computation

The City University of New York

Tapering Detection • Linear Tapering Structure Detection – Tapering to point/line/offset

• Two-step Flow Linear Tapering Detection – Locating potential tapering structure – Examine the converged slice

The City University of New York

Window and Door Installation • Project Windows/Doors on Walls – Identify the closest plane for projection – Push-pull operation on projected vertices

The City University of New York

Experimental Results Cooper Union Model

Experimental Results Spitak Church Model

Experimental Results Thomas Hunter Building

Experimental Results Opernhaus Hannover (Courtesy of the Institute of Cartography and Geoinformatics, University of Hannover)

The City University of New York

Performance Evaluation • Parameter and Error Estimation – Two key parameters – Error measurement

• Limitations – Error propagation – Intersection of two parts

The City University of New York

Parameter and Error Estimation • Key Parameters  d , r 1 • Error Measurement E

X

d

2

( x, M )

xX

The City University of New York

Limitations • Error Propagation • Intersection of Two Parts

The City University of New York

Thank you !