Abstract: Convolutional Neural Networks are the most popular approach to performing image recognition. But how can we move them from the lab to the real world? In this talk Daniel will discuss the challenges of classifying pedestrian demographics in unconstrained environments and using the latest advances in computer vision to solve critical business problems. You can expect to hear about novel image labelling techniques, why people are so valuable, and the future of computer vision.
Bio: Daniel has just completed his PhD in Computer Science and Biometric Identification from the University of Southampton and is now the co-founder and CEO of Aura Vision Labs, a video AI platform specialising in measuring and improving retail shopping experiences. His research involves robust estimation of pedestrian demographics from CCTV imagery using the latest techniques in computer vision and psychological crowdsourcing. Daniel’s research is published in the leading applied machine learning journal, IEEE TPAMI and Aura Vision was featured on BBC Click in May 2018.
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
TensorFlow London 18: Dr Daniel Martinho-Corbishley, From science to startups with Tensorflow, Computer Vision and people.
1. From Science
to Startups
Dr. Daniel Martinho-Corbishley
daniel@auravisionlabs.com
auravisionlabs.com
with Computer Vision,
Tensorflow & People
2. Video AI platform for retailers
to measure and improve
every shopping experience Daniel Jaime Jonathon
PhD in Computer Vision and Soft Biometrics
Recently published “Super-Fine Attributes with Crowd Prototyping” in IEEE TPAMI
Dr. Daniel Martinho-Corbishley
4. Facial Recognition
- Over 3 days to identify suspects
- Extremely hard to spot faces in crowds
- Internet surveillance traffic growing 7x in the next 3 years
11. - Label images as coordinates in super-fine space.
- Precisely describes multiple, integral concepts.
Male Female
Gender
Uncertainty
ClearObscured
X
X
XX
X
Super-Fine Attributes
- Don’t account for ambiguity or uncertainty.
- Irrelevant and inconsistent labels.
- Poorly generalised classifiers.
Female Female Male Male? ???
Categorical / Binary space
Super-fine space
12. - Crowdsource pairwise similarities between n = 95 subjects.
- Forms a high-dimensional distance matrix:
- Embed with Metric MDS to discover fewer, more salient concepts:
- Cluster with Agglomerative Hierarchical clustering to discover c = 5 prototypes
,
Embedding
Clustering
Prototype cluster
Crowd Prototyping 0 n
⋱
⋱
⋱
n ⋱
Distance matrix
13. Male
(0.00, 0.32)
Pos. Female
(0.64, 0.00)
Female
(1.00, 0.29)
Obscured
(0.69, 1.00)
Pos. Male
(0.30, 0.63)
Crowd’s perceptual space &
Visual prototypes.
Matching new images
to visual prototypes.
Efficient, large scale super-fine attributes
14. Very young
Quite young
Quite Old
Very Old
Obscured / Can’t See
Super-Fine Age Labels PETA dataset
- Large-scale - 19000 image samples
- Most diverse - 8799 unique subjects
- 108 binary attributes
15. Super-fine +4.02% AUCSuper-fine +8.25% AUC
Ranked Retrieval
Super-Fine vs Conventional
ResNet-152 Gender & Age [Super-fine & Binary]
3 Attributes [Super-fine]
35 Attributes [Binary]
CNN Training
Binary classified
Super-fine regressed
22. Deep, anonymous insights
People counting is a
competitive landscape
100% Anonymous
Fully GDPR compliant
No personal data stored.
Shoppers are never identified.
Footfall
counts
Heat
maps
Product
footfall
Peel-off
rates
Product
engagement
Dwell
maps
Gender Age
23. The Product
1
Your campaign drew
5% more
females aged 16-24
“
2 3
Minimal Installation
Rapidly integrates with
existing CCTV cameras.
Capture unique insights
Intuitive dashboard and API
reports shopper insights
Define impact
Retailers can now measure
their performance and ROI.
”