Actionable insights from aerial imagery
Actionable insights from aerial imagery
Agenda
TensorFlight demo
How does our software work?
How can we tailor our software to your needs?
Who are we and who are our clients?
Market analysis
Agenda
TensorFlight demo
How does our software work?
How can we tailor our software to your needs?
Who are we and who are our clients?
Market analysis
TensorFlight - team and current usersTeam Users
Drone service partners:
Direct users:
Founders
Advisors
Other Subcontractors - map annotators ● Training data set preparation
Robert Kozikowski● Ex-Google and Facebook engineer● Big data and scaling
Zbigniew Wojna● Deep learning Ph.D. at UCL (World records)● Google DeepMind, Brain Street view, Microsoft, Nvidia
Jan Malaszkiewicz● Strategy consultant, Boston Consulting Group
Deep learning experts● Pieter Abbeel, professor at UC Berkley, OpenAI● Kyunghyun Cho, professor at NYU● Marcin Moczulski Ph.D.
More than 1000 users on DroneDeploy
Source: https://blogs.nvidia.com/wp-content/uploads/2014/09/InsideImagenet.png
Zbigniew - The best model on Image Classification ImageNet 2015
Zbigniew - The best model on MS COCO detection challenge 2016
Source: https://adriancolyer.files.wordpress.com/2016/04/imagenet-fig4l.png?w=1132
Agenda
TensorFlight demo
How does our software work?
How can we tailor our software to your needs?
Who are we and who are our clients?
Market analysis
• DJI drones via DroneDeploy auto pilot app
• Your own aerial imagery - aircraft, drone or satellite
• Terravion
• Orthonormalization and tiling via DroneDeploy or OpenDroneMap
• Request processing directly from TensorFlight or via DroneDeploy app
TensorFlight specializes in Computer Vision
Image analysisMaps formattingData collection
• Accurate - State of the art computer vision via deep learning
• Quick - Scalable, distributed deployment in the cloud
• Usable - Analyze results in depth on web frontend
TensorFlight
We are expanding our Computer Vision capabilitiesAvailable today:• Trees: healthy, unhealthy, dead• Livestock• Cars• Parking spots• Houses• Crops• People
Development in progress:• Trees – recognition of tree type through
crowdsourcing our users• Recognition from plane imagery• Area estimation of plantations• Area estimation of rooftops• Area estimation of damaged rooftops
Showcase - Successful app on DroneDeploy platformTensorFlight
• Dead / Healthy tree counter• Histogram of tree sizes• More than 200 000 acres of land mapped
in our database
1000+ users across the World
Agenda
TensorFlight demo
How does our software work?
How can we tailor our software to your needs?
Who are we and who are our clients?
Market analysis
Step 1: Gathering maps
Step 2: Human annotation of examples to create a training data set (10K-100K examples needed) – we have our own subcontractors
• Image detection - rectangle around each object - 1 month ETA• Image segmentation - detailed line around object and the volume - 3 months ETA
Step 3: Using the training data set to adapt our algorithm
Step 4: Processing the production data set
Our algorithm can be adapted to other object categories
Algorithms achieve human level accuracy on the majority of computer vision tasks
How to map an area?
Small/Medium Big Huge• Purchase a DJI drone• Take pictures, convert to
map format
• Find a company operating a fleet of drones
• Payment method: price per acre
• Example: find a provider on DroneDeploy Mapping Directory
• Use Terravion
• Dependent on needed level of detail
• Consider satellite imagery (price effective, but potentially not enough detail): Planet Labs
Agenda
TensorFlight demo
How does our software work?
How can we tailor our software to your needs?
Who are we and who are our clients?
Market analysis
Reference point – total market size estimates2020 addressable market
Conservative estimate for orchard-related product market is $90m in 2020
OtherInsurance
7
TotalAgriculture
World total permanent crops (orchards, vineyards) land area: ~470m acres1
Innovation adoption rate follows typical Bass diffusion model2, conservative scenario: each acre covered only once
1000
400500
300200
90
Cumulative coverage of permanent crops (m acres)
Most applicable estimate for drone based solutions market size ambition: what is the market value of services that can be replaced by drones by 2020?
1. Source: Worldbank; 2. Standard forecast for adoption rate of a new technology superior to current products; 3. Source: PwC “Clarity from above” report
$B
Conservative scenario: we only play in Insurance area ($7B), not counting Agriculture market
Total addressable market by 2020: $90m, assuming: $1 drone plant-counting related revenue per acreWorldwide sales access and wide coverage of plant recognition models
This is a very conservative estimate – only for permanent crops, doesn’t consider multiple usage on the same area
Q&A