Civil Maps is a company that uses deep learning to automatically generate maps from 3D point cloud data collected via LiDAR. The company's CEO, Sravana Puttagunta, explains how they train neural networks on feature primitives extracted from point clouds to recognize common map elements like poles, wires, and road centerlines. Their system generates millions of algorithm combinations daily to improve its ability to interpret point cloud data and produce accurate maps with minimal human input. Civil Maps stores maps and point cloud data in the cloud and offers visualization and processing tools to customers.

1. SRAVAN PUTTAGUNTA
CEO, CIVIL MAPS
HTTP://CIVILMAPS.COM
A Scalable Approach to Point Cloud
Processing with Deep Learning

3. What and who is behind Civil Maps?
$2.4 Billion $3.5 Billion $3.1 Billion
Our Office
National Energy Research
Scientific Computing Center.
Fastest super computer in
the year 2000

4. Our Mission
Teach computers to make maps
from 3D point cloud data.

5. Point Cloud Summary
 A point cloud file is the output collected from a
complete LiDAR system
 LiDAR Sensor (laser scanning)
 Inertial Measurement Unit (geo-positioning)
 Hard Drive (storage of data)
 Popular formats
 PTS
 LAS/LAZ
 Point Cloud Data

6. Intro to Deep Learning
 Deep Learning is a form of machine learning where
researchers train a computer to find patterns

7. Thought leaders in Deep Learning

8. Map using Deep Learning
A B C

9. 4 Example Feature Primitives for A
 Corner detection
 X distribution
 Y distribution
 Z distribution
 Z-frequency analysis
 Reflectivity analysis
 Spatial consistency

10. 4 Example Feature Primitives for B
 Corner detection
 X distribution
 Y distribution
 Z distribution
 Z-frequency analysis
 Reflectivity analysis
 Spatial consistency

11. 4 Example Feature Primitives for C
 Corner detection
 X distribution
 Y distribution
 Z distribution
 Z-frequency analysis
 Reflectivity analysis
 Spatial consistency

12. Train a Neural Network
1) Filter by
Height
2) Filter by
Reflectivity
3) Find the
corners
4) Check
spatial
consistency

13. Deep Learning = Human Contextualization
HumansComputers

14. Querying with a Neural Network
Query w/ Neural
Network Result Set

15. Deep Learning Demo
 Mark the centerlines every 2 meters
 Find all the poles
 Mark the overhead electrical wires every
meters

16. Deep Learning Demo

17. Manual vs. Deep Learning (X vs. Y)

18. Manual vs. Deep Learning (X vs. Z)

19. Comparative Analysis
Technology Affordability Accuracy Speed Score
Single Person 5 4 1 10
Multiple Person 1 4 1 6
Automated Software 3 2 2 7
Custom Algorithms 2 2 2 6
Deep Learning 3 5 5 13
Highest score is most feasible

20. The Civil Maps System (Intelligence)
• Database of 500 feature
primitives
• Combinations of features 3-4
layers deep, 2 million
algorithms generated
per day
• Scoring algorithm
• High scores survive
• Low scores blacklisted

21. The Civil Maps System (User Interface)
Civil Maps
Customers
Cloud
Storage
1) Upload Point
Cloud Data
2) Upload sample map
for small segment
3) Deep
Learning
4) Map Layers
Uploaded to
Visualizer
5) Report is ready for export

22. The Ask : Join Us
 Users benefit as Civil Maps becomes smarter
 High volume reduces costs for everyone
 Our customers & partners have an unfair advantage
 Free visualization tools
 Processing fees based on number of assets per km

23. Contact Us
 Email : info@civilmaps.com
 Phone : +1 (510) 698-APIS
 Coupon Code : SPAR2015_USER