"A Practitioner’s Guide to Commercializing Applications of Computer Vision," a Presentation from Firelake Capital Management

Copyright © 2016 Firelake Capital 1

Peter Shannon

May 2, 2016

A Practitioner’s Guide to Commercializing

Applications of Computer Vision


Joined Firelake Capital in 2010

• Eye Response Technology startup co-founder

• Systems engineering and computer science by training

• Previously at Atlas Venture from 2005

• Chicago Booth MBA

Investment focus

• Transportation systems

• Food systems

• Energy systems

Peter Shannon

Intersection of

digital and

physical worlds

Copyright © 2016 Firelake Capital 3 3

IT Drivers Big Data

Sensors

Cloud Infrastructure

Computer Vision

Robotics

Sustainability Drivers Distr. Generation (solar)

Batteries

Electrification Complex

Fuel Cells

Ag Innovations

Tech Platforms Industrial IoT

Drones

Autonomous Vehicles

EV Infrastructure

Microgrids

New Bus. Models Software-driven

Finance-driven

Bus. process-driven

“Applications &

Solutions”

Focus on Applications and Solutions


In my earlier days




Key Challenges — Retrospective

Underlying Technology • USB still 1-2 years away

• Cameras bulky, expensive, low resolution

• Facial recognition algorithms still in academia

• Compute power limited (Intel Pentium)

• No computer vision software libraries

Performance vs. Robustness • Each application has different reqs

• Hard to tell when the product is

“robust enough”

Architecture Choices • Hardware

• Balance competing factors

• Upfront choices you must live with


Inductive reasoning – is reasoning in which the premises are viewed as supplying strong

evidence for the truth of the conclusion.

While the conclusion of a deductive argument is certain, the truth of the conclusion of

an inductive argument is probable, based upon evidence given.

Computer Vision is Inductive Reasoning


Deductive vs. Inductive — Development Process

Requirements

Develop

Test

Deploy

Requirements

Develop

Test

Validate Robustness

Deploy


Consequence of Error







Comparison to Speech Recognition




1995 1995


2003 2003


2010 2010


2016 Today


More types of risk in a venture multiplies the overall risk and uncertainty in

terms of time, people, and capital required to achieve the end goal.

Overall Risk Factors

Performance

Robustness

Market

Product

Execution


Implementation Considerations


Decisions you make now are ones you have to live with for the product’s

lifecycle

Conventional wisdom of Minimum Viable Product, speed to market and

failing fast does not apply in the conventional sense

Plan for the End Game


• Need to account for a very large number of factors

• For many applications, a machine-learned algorithm is not sufficient

• Additional application-specific algorithms must be engineered

Expect to spend 10% of time on main algorithm design, 90% on additional

logic for robustness

Achieving Robustness is Difficult


• Hardware-aware software design choices are often necessary

• Picking the best algorithmic approach and the right hardware is difficult,

but a crucial decision that will shape the lifecycle of the product

Vision Products are Hardware-Software Systems


• Your product is a carefully tuned system unifying software and hardware

working together

• What happens when suppliers upgrade components, e.g. cameras, etc.?

• Cannot be certain system performance will automatically improve

• Your application may need to be re-tuned to accommodate the hardware

Supporting The Product Going Forward

This is an ongoing task you need to

plan for and architect your system for


• Q&A

Thank You