Vip Model Call Girls (Delhi) Karol Bagh 9711199171✔️Body to body massage wit...
Big Data LDN 2017: Deep Learning, DeepAd Car Recognition Project
1. Deep Learning:
DeepAd Car Recognition Project
Neil Stobart
Global System Engineering Director
Cloudian
1
2. Machine Learning – Training AI
AI has the power to push ideas and actions beyond human biological limits,
providing us with capabilities to solve problems that were previously too
strenuous or tedious for humans.
ML is a large component of AI that gives
"computers the ability to learn without being explicitly programmed,“
AI pioneer Arthur Samuel, IBM, 1959.
The foundation of AI and ML are data management systems that organize
vast amounts of training data, the essential ingredient for all machine learning
and intelligence.
3. The “DeepAd project” – Making Big Data Smart Data
The “DeepAd project” delivered a digital billboard dynamic content system in Tokyo. DeepAd used
Artificial Intelligence combined with the Internet of Things and Big Data to detect and identify cars—with
94 percent accuracy—on one of Japan’s busy expressways. The system then selected and displayed
content based on the types of cars.
Project Goals
Create an out of home (OOH) advertising platform to target audiences with specific content just for them
Similar to highly targeted ads we experience while surfing the web via a billboard platform
Benefits advertisers with more effective content delivery
Connecting with Consumers through Effective Content Targeting
Improve the effectiveness of billboard outreach.
Big roadside signs and other OOH media do have an impact on behaviour, according to a 2013 Arbitron study
75% of adults notice OOH ads in a month
Over 25% of viewers visit a store, business, or restaurant immediately after seeing an OOH advertisement.
4. Camera installed under a billboard to live-stream
vehicles on Tokyo Metropolitan Expressway
6. Targeted Audiences
The system was trained to recognize specific car models, which
were classified in three different categories for the purpose of
delivering targeted messages:
1. Luxury cars—including all models of Mercedes, BMW, Audi,
and Lexus
2. Family cars—Toyota Prius*, Aqua*, and Vitz,* plus Honda Fit*
3. Project member cars—a 2001 Honda Odyssey*, a 2010
Subaru Outback* BR9, and a 2001 Toyota Bb*.
Control - These cars, driven by members of the DeepAd project passed
by the camera several times
No car detected
Current weather data
Luxury car
Golfing resort
Family car drivers
Local amusement park
Project members
Unique graphic image
7. Training Material
5,000 images of each targeted car type to train the
algorithm on the automobiles the system was to
target.
Nothing special about the images; they were
acquired publicly from automobile manufacturers’
and dealerships’ websites.
The algorithm looked at key features of each car,
such as fender radii (corner angle), headlights, and
other characteristics, to determine the automobile
maker, model, and year.
8. Smart Data Example
Car make, model, year,
and its view angle is
recognized and classified
by S3 key
9. Smarter Storage
One of the reasons why this technology is possible is
through the use of metadata. Typically, big data is just
stored passively for future analysis.
Because this data is unorganized and untagged, it
requires a good amount of effort in order to discover
and pull out specific information.
Object storage allows metadata tags attached to data objects.
As data runs through real-time classification and auto-
recognition/discrimination, metadata tags are attached on the fly.
As a result, we use this ‘deep learning’ to turn big data into smart
data.
11. Sensor
(Video Camera)
Automobiles,
robots,
manufacturing
machines,
monitors …
DL execution
environment
Neural
Model
Sensor Data
(photographed
images)
Real time processing
Recog-
nition Decision Control
CloudIoT / On-site
DL Learning
Environment
Training data
(big volume of
sample images)
(Video Data)
Control Signal
/ Feedback
Neural Model Generation
Edge
Sample Data
DevelopmentExecution / Operation
Sensor Data
Collaboration w/ DL Solution Companies
Deep Learning + HyperStore
✓ HyperStore enables meta data management, which is a good fit as storage for a DL
development environment where labeling of big volume of training data is required.
12. AI-based automated traffic census
AI and IoT can drive a shift of traffic survey from manual sampling to 24-hour, high-speed
measurement and aggregation
Deep learning-based traffic volume survey
• Automated counting, and drastic reduction of
measurement cost
• 24-hour measurement of traffic volume in details, e.g.,
congestion at multiple locations
• Instantaneous data aggregation and display
13. Measurement of buses and passengers at Shinjuku
Bus Terminal
• Automated AI-based measurement,
instead of data collection from bus
companies
• Conventional approach: Difficult to fully
collect data from 180 or more bus
companies
• Conventional approach: One month or
more to collect data which contains
significant errors and is based on
different measurement standards
• Target: AI-based data aggregation within
the next day and 365-day full automation
• Released on the 1st anniversary of
opening of Shinjuku Bus Terminal
14. What else? Security Cameras..
Road Monitoring
❖ traffic analysis
➢ vehicle counts
➢ pedestrian counts
❖ event detection
➢ falling objects
➢ flooding
Human Monitoring
❖ human behavior analysis
➢ suspicious
■ looking around
■ leaving objects
➢ help needed
■ disabled
■ elder
Why DL so interesting?
Provides very high accuracy in generic
object recognition with ease
NOTE: Good at discriminative object recognition like face
recognition and number plate recognition under strict proved
environments
OUR FOCUS PARTNER
15. Cloudian: AI Data Management
The AI Gold Rush = Massive Data Needs
Industrial IoT
Outdoor advertising
Traffic management
Server failure prediction
Electricity Management
Semiconductor defect detection
Machine parts routing
Movie pixilation
16. HyperStore Use Cases
Storage as a Service
QoS
Billing Multi Tenancy
Clients
S3 Storage
Service
APPLICATIONS BACKUP SERVERS
Backup and Archive
S3
INPUTS PROCESSING PACKAGING STORAGE DISTRIBUTIO
N
Media and Entertainment
Result of
Analysis
Hadoop Map Reduce
Apache Spark on Cloudian platform
Cloudian Platform
Cloudian HyperStore
Analytics
Social Media
Device Tracking
& Logs
Bioinformatics
AI, ML, Analytics