Industrial Imaging Today & Tomorrow AIA 2016 The Vision Show. Steve Varga, Procter & Gamble

Transcription

1 Industrial Imaging Today & Tomorrow AIA 2016 The Vision Show Steve Varga, Procter & Gamble

2 Steve Varga Next big breakthroughs

3 Procter & Gamble Countries of Operations ~70 Countries Where Our Brands Are Sold 180+ Annual Sales $70B+

4 Imaging & Instrumentation Capability R&D 2 to 7 years Off the shelf Technology road-maps

5 R&D Partnerships

6 The Vision Industry today is amazing Servos vs. mechanical drives 20 years ago

7 Paint a Picture of Tomorrow

8 3D vision is maturing Strong history in robotics High speed profilometry is here FPGA Field Programmable Gate Arrays and dedicated GPUs

9 Motion Analysis Many examples in sports and automotive testing 3D motion is improving with multi-camera systems but still requires special markers for most applications but not for long. 3D fluid motion analysis still has a ways to go

10 Learning how we see Dr. G Francis Purdue University Dr. L Itti University of Southern California Dr. B Lotto founder of LottoLab Dr. S Grossberg Boston University

11 Figure-ground segmentation Human visibility and it s relation to computer algorithms Which elements are in front?

12 Figure / Ground Perception Occlusion, Competition, Modal Surfaces, Amodal surfaces In this image you have a modal percept of the vertical lines Black color You also have an amodal percept of a horizontal contour It has no color

13 Amodal percepts Is amodal perception just inference? NO! Statistical inference implies the occluded check should be black But it looks (amodally) to be white Kanizsa, 1979

14 Craik- O Brien- Cornsweet Effect

15 Without the boundary

16

17

18 Feature system Color must fill-in or spread across a defined surface Pinna et al., 2001

19 Lotto: Color is a perception Light is real; color is not Color is a learned interpretation of our senses It is learned, just like language We have learned what shadows do to reflective surfaces

20 We learn what to perceive based on our experience Dr. Beau Lotto

21 Bottom-up Process: We don t think about it

22 Deep Learning Machine Learning: AI rebranded First AI? Arthur Samuel 1949 w2 w1 w3 AI Magazine Volume 11 Number 3 (1990) 1961 His computer beat Robert Nealey, who went on to become the Connecticut State champion in 1966.

23 No War Games; No Skynet 2007 TED (March 2015) How we're teaching computers to understand pictures

24 Deep Learning Girl Teddy Bear A woman is dancing in a field of yellow flowers blanket

25 What s next? Hello, my name is Watson Deep Learning can seem to learn just about anything, so what s next?

26 Kind of scary? Computers are demonstrating that they can learn tasks which we don t know how to code. Computers are learning how to do complex tasks that we ourselves don t know how to do!

27 Deep Learning Lawyers Journalists Surgeons Financial Consultants Marketing / advertisers

28 Robotic Surgery

29 Why haven t we seen it yet? Feature Engineering still facing the inability to efficiently extract and organize the discriminative information from the data.

30 Saliency-Based Analysis

31 Saliency Software at

32 Two-streams hypothesis Object identification Dorsal (Where?) Location: Motion & Figure / Ground segmentation Ventral (What?)

33 NVIDIA GEOINT & DIGITS Hardware and software specifically designed for machine learning. DIGITS free and available for download. NVIDIA DGX-1 GPU designed especially for deep learning.

34 Cognition and Reasoning John Kremer

35 True AI may still be fiction but it s close enough for industry Kim Kyung Hoon/Reuters Robot HRP-2 demonstrates use of a tap after washing a cup at Tokyo University

36 Big Data & Digitization & Deep Learning Why not manufacturing tasks? Data systems now have the information but the factory is still missing its eyes. Images are needed for the machines to learn to see. 4M / sec Absorbent and assembled products Convolution neural networks: Quality Assessments Defect detection Process control Machine health and preventative maintenance Operator feedback Consumer preferences Retroactive consumer feedback for process and machine control

37 To See is to Understand The web is wrinkled and elastic 7 is obscured

38 Hundreds of companies specializing in vision and related products / services

39 Thank you for Kristi

40 Thank you Steve Varga Procter & Gamble

41 Speed (cost) and accuracy

42 3D Advancements In 2011, Lytro released the 1 st consumer lightfield camera. Max Born Emil Wolf Principles of Optics, 1959 High Density GPUs are approaching capabilities to solve complex wavefront equations in real-time. In-line high speed 3D microscopy is coming!