Decoding Kotlin - Your guide to solving the mysterious in Kotlin.pptx
Computer Vision problematics in the industry by Despoina Ioannidou, Team Lead R&D @Meero
1. Computer Vision problematics in the
industry: the Meero case
Despoina Ioannidou
8/10/2019
Paris Women in Machine Learning and Data Science
2. Meero’s mission
On-demand end-to-end photo shoots delivery
Clients
Client orders shoots
& pays Meero
1
Meero matches it with
suited photographers
2
Meero qualifies, trains &
onboards photographers
0
Photographers
shoot
3
Photographers
upload DSLR photos
& are paid by Meero
4
Photographs go
through AI editing &
human final touch
5
Client retrieves
the media
6
Handling 2000+ shoots a week in 90+ countries
3. Where computer vision and DL come in: Scalability
Upgrade a photo shoot to a professional look
And make it automatic…………...
4. Automatic editing - the journey
Summarize all the problems to enhance an image
➔ white balance
➔ exposure
➔ color enhancement
➔ sharpness
➔ contrast enhancement
➔ green-magenta tints
➔ noise
➔ compression artefacts
➔ blur
➔ dirt
➔ low point of view
➔ rotation
➔ tilt
➔ shadows attenuation
➔ flash reflections
➔ lens distorsions
….
1
round
Test automatic software editors...
2
round
3
round
6. AI editing - the journey
focus on the most salient issues: exposure & color enhancement, white balance
Photographer’s shoot Expected result
1
round
2
round
3
round
7. Auto mode of photo editing software
Photographer’s shoot Editor #1: yellow, underexposed
AI editing - the journey 1
round
2
round
3
round
8. Auto mode of photo editing software
Photographer’s shoot Editor #2: yellow & magenta, underexposed
AI editing - the journey 1
round
2
round
3
round
9. Auto mode of photo editing software
Photographer’s shoot Editor #3: too saturated, underexposed
AI editing - the journey 1
round
2
round
3
round
10. second round
feedbacks - Several bias per editor
(green-magenta bias, underexposure, saturation, etc.)
- Global settings are not sufficient to edit a photo with professional
quality result
- Image segmentation to apply locally tuned changes is not a solution,
it takes quite a while for mitigated results
- Our images are HUUUGE (Minimum 15Mpix, up to… 50Mpix)
AI editing - the journey 1
round
but... - We have a very wide dataset of pro-edited real estate images
- We have no processing constraints for production
2
round
2
round
3
round
11. AI editing - the journey 1
round
3
round
2
round
define the
problematic
- What can we build to allow meero work faster?
- What does automatic retouching consists in?
Case of automatic retouching:
- Technical retouching: noise, artifacts, blur,...
- Geometry retouching: rotation, tilt,...
- Aesthetic retouching: colours, exposure,...
break it down in
smaller modules
assess feasibility
- What can be attained in a week, in a month, in a year?
- What can be attained by development, what is research and what
is r&d
12. AI editing - the journey 1
round round
3
round
2
our toolbox
14. … based on
M. Gharbi, J. Chen, J. T. Barron, S. W. Hasinoff, F. Durand
Deep Bilateral Learning for Real-Time Image Enhancement - Siggraph 2017
AI editing - the journey 1
round
2
round
3
round
Design our own
conv-deconv deep
network
15. and on...
O. Ronneberg, P. Fischer, T. Brox,
U-Net: convolutional networks for biomedical image segmentation - CoRR 2015
AI editing - the journey 1
round
2
round
3
round
Design our own
conv-deconv deep
network
16. Auto mode of photo editing software
Photographer’s shoot Editor #1: yellow, underexposed
Reminder-other editors
17. Auto mode of photo editing software
Photographer’s shoot Editor #2: yellow & magenta, underexposed
Reminder-other editors
18. Auto mode of photo editing software
Photographer’s shoot Editor #3: too saturated, underexposed
Reminder-other editors
19. Here are the results...
AI editing - the journey 1
round
2
round
3
round
20. AI editing - the aftermath
Did we achieve our goal? Is the retouched image
beautiful enough?
- Experts opinion is guiding us but...
- Research motivated by this question:
Aesthetic metric learning: Can we teach a machine
to learn to “understand” how aesthetically pleasing
an image is? What features make a photo
21. Are AI edited images still photographs ?
Editing changes the way
you see the photos. It is
like fake news !
AI will have biases that
the conceiver will
maybe not forecast
Photo is an art - emotion is
better than accuracy
I want to feel the same
atmosphere as during the
shoot, editing is important
The truth is the raw
image from my
camera
The cultural background is key,
images will look average
Reality is in the
eye of the
observer
22. And it is only the beginning...
Scene parsing and understanding
Unified Perceptual Parsing for Scene Understanding, T Xiao, 2018
23. ...of a new love story
It’s a new match ;)
AI Camera
born in 1956 born in 1827