SLIDE 1
Situational Awareness for Smart City: Opportunities and Challenges
Hao Lu | hao.lv@yitu-inc.com
About YITU AI
Health
Finance
Public Safety
Retail
Situational Awareness for Smart City: Opportunities and Challenges - - PowerPoint PPT Presentation
Situational Awareness for Smart City: Opportunities and Challenges - - PowerPoint PPT Presentation
Situational Awareness for Smart City: Opportunities and Challenges Hao Lu | hao.lv@yitu-inc.com About YITU AI Public Safety Retail Health Finance Face Recognition Price Challenge 2017 Winner In Face Identification
SLIDE 2
SLIDE 3
Face Recognition Price Challenge 2017
Winner
In Face Identification
SLIDE 4
<1 billionth false match rate
SLIDE 5
<1 billionth false match rate But, what does it mean?
SLIDE 6
0.20 0.15 0.10 0.05 0.00 30 40 50 60 70 80 90 100 0.20
score percentage
SLIDE 7
0.20 0.15 0.10 0.05 0.00 30 40 50 60 70 80 90 100 0.20
85 95 90 66 75 60 40 1/10M 1/1B 1/100M 1/100k 1/1M 1/10k 1/1k False Alarm Rate Similarity Score
score percentage
SLIDE 8
0.20 0.15 0.10 0.05 0.00 30 40 50 60 70 80 90 100 0.20
85 95 90 66 75 60 40 1/10M 1/1B 1/100M 1/100k 1/1M 1/10k 1/1k False Alarm Rate Similarity Score
Twins
score percentage
SLIDE 9
0.20 0.15 0.10 0.05 0.00 30 40 50 60 70 80 90 100 0.20
85 95 90 66 75 60 40 1/10M 1/1B 1/100M 1/100k 1/1M 1/10k 1/1k False Alarm Rate Similarity Score
score percentage
SLIDE 10
0.20 0.15 0.10 0.05 0.00 30 40 50 60 70 80 90 100 0.20
85 95 90 66 75 60 40 1/10M 1/1B 1/100M 1/100k 1/1M 1/10k 1/1k False Alarm Rate Similarity Score
score percentage
Cost
SLIDE 11
Towards Situational Awareness
Opportunity: City can better understand how infrastructure is serving citizens
SLIDE 12
Towards Situational Awareness
Opportunity: Shops can have a deeper understanding about their customers
SLIDE 13
One common theme: re-identify a person
SLIDE 14
One common theme: re-identify a person
Video ID
SLIDE 15
Re-identify a person
Cluster Identify
Across cameras; Across time; At large scale
Capture
SLIDE 16
City Scale
10 million people 100k cameras
SLIDE 17
City Scale
10 million people 100k cameras 1 billion faces per day
SLIDE 18
Capture faces
Detection Tracking Alignment Quality Filter
SSD w/ VGG VGG VGG
SLIDE 19
Capture faces
Detection Tracking Alignment Quality Filter
SSD w/ VGG VGG VGG
SLIDE 20
Face Quality Data
Detection Tracking Filter
100k hours of videos
1:1
Good and Bad Frames Tracks of faces
SLIDE 21
Capture faces
Video streams from 100k cameras
=
lot of bandwidth!
SLIDE 22
Capture faces on the camera end
Feature/headshot batch
TX2 Box
+
With capture
SLIDE 23
Cluster the same faces
- Hierarchical Clustering on CPU
- With heuristics: e.g., cluster same camera first
SLIDE 24
Identify against clusters
- ANN on GPU
- Using a few thousands of Tesla P4
SLIDE 25
Early results - Performance
SLIDE 26
Early results - accuracy
- #Clusters to #people: 3 to 1
- Long tail
SLIDE 27
Discussion
- Camera angles can be different
- Poses can be different
SLIDE 28
Going forward
- Leverage more attributes, e.g., wearing
glasses, same outwear
- Leverage history data
SLIDE 29