2. 2
If you buy a car that does not have the
hardware necessary for self driving, it is like
you are buying a horse…
Google is working on self-driving cars, and
they seem to work. People are so bad at
driving cars that computers don't have to be
that good to be much better.
3. 3
Tesla: 8 Cameras
Mobileye: 13 Cameras
Waymo: 29 Cameras
Tesla Semi Truck: 28 Cameras
What are the challenges to Camera based AV’s ?
5. Challenging condition: Light transition
Frames before, during and after the
transition
5
T = 0.22s
Under exposed
T = 1.2s
Over exposed
T = 1.7s
Correct
7. 7
They have to be a LOT better!
If Autonomous vehicles are relying on images
from cameras, to make driving decisions, those
images need to be of the highest quality, at all
times in all conditions!
Remember this Quote:
Google is working on self-driving cars, and
they seem to work. People are so bad at
driving cars that computers don't have to be
that good to be much better.
How can we get there? By Working together!
8. IEEE-P2020
8
IEEE-SA P2020 working group on automotive imaging has been preparing new
standards and performance metrics (KPI’s) to reduce the ambiguity in
measurement of image quality of human vision and computer vision based
cameras vis-a-vis their intended purpose.
9. DXOMARK Image Labs - An international team,
Led by experts in Imaging, Audio, Display & Battery
Testing and Development Services
9
2200+
m2 office space
16
labs
100+
experts
X4
In 3 years
11. Our history
For more than 10 years, DXOMARK has been setting the industry's standards for image quality and working with all camera and
smartphone vendors to spur them to innovate and develop products with optimized quality for the benefit of end users.
2003
Launch of Analyzer
2008
Launch of
DXOMARK.COM with
DSC testing
2012
Smartphone camera
testing
2017
DXOMARK becomes
an independent
company
2016
100th Analyzer lab
installed
2019
Smartphone selfie &
Audio testing
2020
Smartphone Display &
Wireless Speakers Testing
11
2021
Battery Testing
2022
Videoconferencing
Security,
Automotive, Drone,
360/VR cameras
testing
12. Comprehensive protocol development adapted to the
applications
Applications
Objective Tests LAB
Perceptual Tests
REAL-LIFE
Use Cases
WDR
LED Flicker
Night Vision
Moving Object
Flare
Image Quality Evaluation
report
12
13. In-cabin camera – human vision and DMS
Portrait Color Timing Setup HDR Portrait Setup
Face exposure convergence
White Blance and Color stability
Color fidelity
Frame rate, shutter time, rolling shutter
Face details preservation
13
Perceptual field tests
HDR Face exposure
Entropy on bright pannel
HDR Face detail preservation
Objective lab tests
Spatial visual video noise
Temporal visual video noise
Exposure
Detail preservation
Noise
Focus range
Video visual noise chart DXOMARK Chart
Focus range chart
14. Front facing camera evaluation
14
Perceptual field tests Objective lab tests
15. Services for camera development
15
Support in tuning phase
Specifications for camera image quality
Measurement services and full
report delivery
Camera benchmark
17. T H A N K Y O U !
24-26, quai Alphonse le Gallo - 92100 Boulogne-Billancourt - France
www.dxomark.com
Editor's Notes
Camera is the eye of an autonomous vehicle
If a camera miss an information, no matter how good is the AI, it might affect the AV decision and impact safety
If you choose to use camera, it must be as good as possible
What are the challenges?
High dynamic scene
This is a typical scene
No information in the bright part. If traffic light, obstacle after at the end of the tunnel, at this point, the car is totally unaware and blind
Same thing when we enter the tunnel
= dangerous
Sudden light transition: Use case: when we enter a tunnel
Scene recorded in one of our colleagues car >> when we did that, it created an unwanted emergency break on his adas system
Issue with the adaptation of the auto exposure
Image goes under exposed to over exposed before reaching a correct image
More and more LED in automotive environment on cars, traffic lights, lightings.
The LED lighting is not continuous but our eyes do not perceive the blinking
The camera video acquisition is sampled, that cause this Flicker effect
Car lighting can easily be confused with turning light, traffic light might not be detected at all in some frames >>> might lead autonomous vehicle to take the wrong decision
= danger
Point out OEMs, Tier one, Autonomous vehicle platform makers, Chipset, sensor makers
The P2020 Committee goal is to define Camera Image Quality Standards, both human vision and Computer vision
Who we are
For several years now, we are considered as a reference in camera evaluation
Protocol methodology
A lot of different application for automotive camera from human vision to computer vision application (object detection, line detection, driver monitoring system…)
Use case list for each application
For each use case we have objective tests lab and perceptual test
> this is how we establish a complete report
This is a first example of a protocol we have developed for a specific application.
Goal: - Monitor driver’s behaviors under any condition
Perceptual field test (any lighting conditions, daylight, HDR, night) all skin tones
Description of objective lab tests and metrics. Describe PTC with dynamic lightings and HDR portrait.
It’s important to have a face in the lab setup to fit real use case.
Other example: Front facing camera
Need to visualize and detect object long distance
Description of perceptual field tests and objective lab tests. Emphasize on Collimator + P2020 metrics: Flicker, Flare, Noise, Dynamic range
We can work in all the camera development phase.
We can help OEM’s give specifications to their suppliers
We can help camera makers in their optimization phase
We can benchmark cameras for specific applications.