Image-based Airborne LiDAR Point Cloud Encoding for 3D Building Model Retrieval

Image–based Airborne LiDAR

Point Cloud Encoding for

3-D Building Model Retrieval

Yi–Chen Chen

ISPRS 23

rd

1. Created manually

2. Generated by LiDAR or photogrammetry

3. Reuse data from Internet

Basic Idea

•

Data/Model reuse

• An efficient tool to acquire 3-D spatial data

3-D Model Retrieval System

• For Automatic city modeling

Reuse

Reconstruction

Input

Retrieval System

What's the challenge?

Past

Now

• Develop an encoding approach for

• Point cloud

• 3-D model

• Similarity = Comparison of encoding results

Related work

• Existing studies of model retrieval

• Model–based

• View–based

• Point cloud as input

Spherical Harmonics

Funkhouser et al., 2003

Chen et al., 2003

Point Cloud Encoding using Spherical Harmonics

Chen et al., 2014

System Workflow

Database

Query

Query

Result

3-D Models

Point Cloud

(Input)

Image–

based

Encoding

Why to encode

Image–based Encoding

Building

Roof

Description

Geometric

Features

Spatial

Histogram

Encoding – Depth Image

• Motivation

• Less information in side parts

• More information in roof parts

Top View

Encoding – Depth Image

Max. Height

Depth Image

Building Object

Encoding – Geometric Features

• Height feature

• Height from the ground

• Line feature

• Laplacian of Gaussian edge detection

• Eigen feature

• Planarity

Encoding – Geometric Features

Height

Line

Eigen

3–D

Model

Point

Cloud

1 2 3 4 5 6 7 8 9 10 Hi s togr a m

Encoding – Spatial Histogram

Origin Determination

Center of mass in 3–D

Rotation and translation Invariance

Center of Mass in 3–D

Center of Minimal Bounding Box

Height Feature

Line Feature

Eigen Feature

Encoding – Spatial Histogram

1 2 3 4 5 6 7 8 9 10 P er c en ta g e Bin 1 2 3 4 5 6 7 8 9 10 P er c en ta g e Bin 1 2 3 4 5 6 7 8 9 10 P er c en ta g e Bin

Encoding Property 1

• Consistent encoding of point clouds and

building models

Height

Line

Eigen

Encoding Property 2

• Demonstration of rotation invariance

0°

Encoding Property 3

• Demonstration of noise insensitivity

Point Cloud

Evaluation

1. Encoding of various level-of-detail (LoD) models

2. Data retrieval from a database

LoD – Experimental Result 1

Point Cloud

Detail-1: 1.024

Detail-2: 0.905

LoD – Experimental Result 2

Comparison – Data Retrieval

• Point cloud encoding for 3-D building model

retrieval

• Model–based

• Spherical harmonic function

Reference:

Chen, J.–Y., Lin, C.–H., Hsu, P.–C., and Chen, C.–H., 2014. Point cloud

encoding for 3-D building model retrieval. IEEE Trans. on Multimedia, 16(2),

pp. 337–345.

Database – Experimental Result 1

The Proposed Method

0.38 1.87 2.21

2.21 3.48

3.82 _2.44 1.91

2.50

2.08

The Compared Method

0.38 2.15 _2.07 2.38 2.23

Building #1

Val. = RMSE

Database – Experimental Result 2

Building #2

The Proposed Method

2.16 7.17 8.40 5.82 7.60

5.96 5.96 5.96 6.04 7.17

The Compared Method

2.16 5.84 5.21 8.88 8.52

7.64 7.00 9.34 5.14 10.42

Val. = RMSE

Database – Experimental Result 3

Building #3

The Proposed Method

3.60 2.07 2.07 1.97 1.90

3.87 3.87 4.08 4.08 3.10

The Compared Method

4.94 4.94 2.26 4.08 5.65

Val. = RMSE

Database – Experimental Summary

The Proposed Method

Building 1

Building 2

Building 3

RMSE Avg.

1.89

6.22

3.06

Rank Diff. Sum

50

32

24

The Compared Method

Building 1

Building 2

Building 3

RMSE Avg.

2.29

7.01

5.10

Rank Diff. Sum

63

64

83

(

)

(

)

Conclusions and Future Work

• 3-D Building models in the database and the

input point cloud can be consistently and

accurately encoded.

• Spatial histogram introduces the properties of

rotation invariance and noise insensitivity.

• A web–based model retrieval system has been

developed.

• Description of geometric shape and building

size are added for whole building.