1
First experiments of deep learning on LiDAR point clouds for classification of urban objects Y. Zegaoui¹, M. Chaumont¹², G. Subsol¹, P. Borianne³, M. Derras ¹LIRMM, CNRS/University of Montpellier, France, ²University of Nîmes, France, ³AMAP,Univ. Montpellier, CIRAD, CNRS, INRA, IRD, Montpellier, France, Berger-Levrault company, Toulouse, France [email protected] INTRODUCTION Urban environment monitoring and management [5]. Dynamic LiDAR acquisition to scan entire scene. Urban object detection and recognition. Projecting all collected data in GIS. Deep-learning [4] methodology for shape classification. 3D URBAN OBJECT DATASETS 1 urban object = 1 point cloud Subsampling the point clouds to 512 points From public datasets: (727 objects) Sydney urban dataset: http://www-personal.acfr.usyd.edu.au/a.quadros/objects4.html Kevin Lai dataset: https://sites.google.com/site/kevinlai726/datasets Paris rue Madame dataset: http://cmm.ensmp.fr/~serna/rueMadameDataset.html DEEP-LEARNING FOR 3D POINT CLOUD CLASSIFICATION Voxelizing the clouds [3] Using multi-views [2] Learning directly on point [1] SELECTED METHOD: POINT-NET [1] Points (x, y, z) are directly processed Coordinate frame normalized with T-Net Invariant to order of points CLASSIFICATION EXPERIMENTS Test set: our point clouds PointNet network Network predictions Training set: urban objects from 3 public datasets CONCLUSION Encouraging results for classification Class confusion “traffic sign”/”pole” Limited size of the dataset FUTURE WORK Acquisitions of new datasets Object localization in a scene 3D+t analysis of scene variation Tree Car Traffic sign Person Pole Our dynamic LiDAR acquisition: 200 meters back and forth with Leica Pegasus backpack 27 millions points (x, y, z) + RGB 174 urban objects isolated manually REFERENCES [1] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. C. R. Qi, H. Su, K. Mo, L. Guibas. CVPR, 2017 [2] Synthesizing 3D Shapes via Modeling Multi-View Depth Maps and Silhouettes with Deep Generative Networks. A. Arsalan et al. CVPR, 2017 [3] OctNet: Learning Deep 3D Representations at High Resolutions Gernot Riegler. A. Osman, A. Geiger CVPR, 2017 [4] Deep learning. Y. Lecun, Y. Bengio, G. Hinton. Nature, 2015 [5] Smart City Concept: What It Is and What It Should Be. I. Zubizarreta, A. Seravalli, S. Arrizabalaga. Journal of Urban Planning and Development, 2015 We thank Leica Geosystems for its technical support as well as providing the LiDAR acquisition. This work is supported by a CIFRE grant (ANRT) image from : Geo-Plus VisionLidar final point features 512x1024 Max pooling global features 1x1024 Classifier classes score 1xK input points cloud 512x3 aligned points 512x3 MLP (64, 64) point features 512x64 T-Net T-Net aligned point features 512x64 MLP (64, 64) Mobile LiDAR devices : car platform backpack

First experiments of deep learning on LiDAR point clouds for …chaumont/publications/SFPT-2018_ZEGAOUI... · 2018. 5. 31. · [4] Deep learning. Y. Lecun, Y. Bengio, G. Hinton. Nature,

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: First experiments of deep learning on LiDAR point clouds for …chaumont/publications/SFPT-2018_ZEGAOUI... · 2018. 5. 31. · [4] Deep learning. Y. Lecun, Y. Bengio, G. Hinton. Nature,

First experiments of deep learning on LiDAR point clouds for classification of urban objects

Y. Zegaoui¹⁴, M. Chaumont¹², G. Subsol¹, P. Borianne³, M. Derras⁴¹LIRMM, CNRS/University of Montpellier, France, ²University of Nîmes, France,

³AMAP,Univ. Montpellier, CIRAD, CNRS, INRA, IRD, Montpellier, France, ⁴Berger-Levrault company, Toulouse, France

[email protected]

INTRODUCTION

● Urban environment monitoring and management [5].● Dynamic LiDAR acquisition to scan entire scene. ● Urban object detection and recognition.● Projecting all collected data in GIS.● Deep-learning [4] methodology for shape classification.

3D URBAN OBJECT DATASETS

● 1 urban object = 1 point cloud● Subsampling the point clouds to 512 points

● From public datasets: (727 objects) ○ Sydney urban dataset:

http://www-personal.acfr.usyd.edu.au/a.quadros/objects4.html ○ Kevin Lai dataset:

https://sites.google.com/site/kevinlai726/datasets ○ Paris rue Madame dataset:

http://cmm.ensmp.fr/~serna/rueMadameDataset.html

DEEP-LEARNING FOR 3D POINT CLOUD CLASSIFICATION

● Voxelizing the clouds [3]

● Using multi-views [2]

● Learning directly on point [1]

SELECTED METHOD: POINT-NET [1] ● Points (x, y, z) are directly processed● Coordinate frame normalized with T-Net● Invariant to order of points

CLASSIFICATION EXPERIMENTS

Test set: our point clouds

PointNet network

Network predictions

Training set: urban objects from 3 public

datasets

CONCLUSION

● Encouraging results for classification● Class confusion “traffic sign”/”pole”● Limited size of the dataset

FUTURE WORK

● Acquisitions of new datasets● Object localization in a scene● 3D+t analysis of scene variation

Tree Car Traffic sign Person Pole

● Our dynamic LiDAR acquisition: ○ 200 meters back and forth with Leica Pegasus backpack○ 27 millions points (x, y, z) + RGB ○ 174 urban objects isolated manually

REFERENCES

● [1] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. C. R. Qi, H. Su, K. Mo, L. Guibas. CVPR, 2017

● [2] Synthesizing 3D Shapes via Modeling Multi-View Depth Maps and Silhouettes with Deep Generative Networks. A. Arsalan et al. CVPR, 2017

● [3] OctNet: Learning Deep 3D Representations at High Resolutions Gernot Riegler. A. Osman, A. Geiger CVPR, 2017

● [4] Deep learning. Y. Lecun, Y. Bengio, G. Hinton. Nature, 2015● [5] Smart City Concept: What It Is and What It Should Be. I. Zubizarreta, A.

Seravalli, S. Arrizabalaga. Journal of Urban Planning and Development, 2015

We thank Leica Geosystems for its technical support as well as providing the LiDAR acquisition. This work is supported by a CIFRE grant (ANRT)

Ground Truth

Tree (75) Car (69) Traffic sign/light (8) Pole (23) Person (15)

tree 69 2 0 8 0

car 1 33 0 0 0

traffic sign/light 4 0 4 12 2

pole 0 0 3 1 0

person 1 0 1 00 12

building 0 0 0 02 0

noise 0 4 0 000 1

F measure 0.896 0.904 0.267 0.074 0.828

image from : Geo-Plus VisionLidar

final

poi

nt

feat

ures

512x

1024

Max

poo

ling

global features1x1024 C

lass

ifier

classes score 1xK

inpu

t poi

nts

clou

d 51

2x3

alig

ned

poin

ts

512x

3 MLP

(6

4, 6

4)

poin

t fea

ture

s 51

2x64

T-N

et

T-N

et

alig

ned

poin

t fe

atur

es51

2x64 M

LP

(64,

64)

Mobile LiDAR devices : car platform backpack