TensorFlight object recognition catalog

Table of Contents

1 Introduction

TensorFlight automatically analyzes satellite, aerial, drone and street view imagery in order to automate slow and expensive manual in-person property inspections.

We are primarily focused on commercial property insurance, but some of our automatic recognition capabilities can be also useful in other industries.

If you are a property insurer TensorFlight can help you with:

  • Underwriting: Automatically pre-fill a quote - e.g. square footage of the roof. Highlight areas to investigate for an underwriter - e.g. potential building degradation.
  • Reinsurance: Get more detailed information about each property in a reinsurance policy or gain alpha while trading cat bonds.
  • Risk: Better understand exposure of your portfolio or monitor transient risks.
  • Claims: Automatically resolve simple claims - e.g. a few tiles missing from the roof. Prioritize adjuster activity after a catastrophe like a hurricane.

Simply let us know what properties are you interested in - for more details see the section How to tell TensorFlight what do I want to analyze?.

2 Why partner with TensorFlight?

2.1 Past expertise at leading companies

We worked for multiple years on practical machine learning, big data and GIS projects at companies like Google, Facebook or DeepMind, including transcribing Street View imagery for Google Maps, ranking insurance quotes for Google Compare or processing petabytes of data for Google Search.

This expertise allows us to build reliable, scalable and performant system ensuring our data will be always delivered to you on time.

2.2 Computer vision world records

Zbigniew is co-author of Inception v3, famous deep learning neural network structure and other influential computer vision research. Inception v3 was the first computer vision model that beat human accuracy in image classification achieving 96% world record accuracy on ImageNet in 2015. ImageNet is the most important computer vision benchmark, classifying 1M images into 1000 classes.

This proves that we can build the most accurate automated imagery analytics.

2.3 Necessary domain expertise and focus on commercial property insurance

In addition to computer vision experts, our team consists of civil engineering post doc and ex-insurance expert. What's more, TensorFlight is the only computer vision company focusing on commercial property insurance.

This ensures that our analytics will be relevant and easy to consume for companies in P&C insurance market.

3 Objects supported by TensorFlight

3.1 Current capabilities

Object type Output type Explanation Example Average model accuracy
Commercial building footprint Polygon Footprint of commercial or industrial building - e.g. shopping mall or gas station. commercial_catalogue.jpg 80%
Residential building footprint Polygon Footprint of residential building - either house or block of flats. residential_catalogue.jpg 80%
Temporary building footprint Polygon Building-like structure that is not permanent - e.g. tent or trailer. temporary_catalogue.jpg 80%
Missing part of building Polygon Part of the building that is missing a standard structure - e.g. it is still in construction or part of the roof that have been torn by a Hurricane incomplete_catalogue.jpg 80%
Building degradation Polygon Property degradation that is not affecting a structural integrity of the property - e.g. pooling water, missing shingles, facade paint cracking. degradation_catalogue.jpg 80%
Tree Bounding box Tree - posing risk of catching fire or falling under heavy wind. tree_catalogue.jpg 95%
Leafless tree Bounding box Leafless tree is likely dead, so it's posing higher risk of fire and wind than tree with leaves. dead_tree_catalogue.jpg 95%
Vehicle Bounding box Vehicle - e.g. car, truck or ship. vehicle_catalogue.jpg 95%
Parking space Bounding box Parking space - counting empty parking spots helps estimate activity around the property. parking_catalogue.jpg 95%
Solar panels Polygon Group of solar panels. If multiple solar panels are adjacent they are counted as a one group. solar_panel_catalogue.jpg 80%
Wind-borne debris Bounding box Items nearby property that can be blown by the wind and damage the property envelope - e.g. chairs, lumber, piles of trash. debris_catalogue.jpg 80%
Mechanical equipment on roof Bounding box Mechanical equipment on the roof, that can be blown by the wind - e.g. HVAC. mechanical_equipment_catalogue.jpg 80%
Pool Polygon Swimming pool. Other water sources like ponds or fountains are not included. pool_catalogue.jpg 80%
Window or skylight Polygon Either skylight on the roof or window on the facade. skylight_catalogue.jpg 80%
Antenna or satellite dishes Bounding box Antennas or satellite dishes on the roof that can be blown by the wind. antena_catalogue.jpg 80%

3.2 Future Road map

Object type Output type Explanation Example Average model accuracy
Construction type Classification Classification of construction type - e.g. wood frame, masonry, tilt-up, engineered structure   ETA 3 months
Number of stories Classification Count of stories in the building.   ETA 3 months
Occupancy type Classification Type of building usage - e.g. Retail trade, Professional technical and business, Apartment/Condo, Restaurants.   ETA 3 months
Building Height Regression Height of building in meters.   ETA 6 months
Estimated building age Regression Estimate building age based on the visual characteristics.   ETA 6 months
Signage Bounding box A sign that poses high risk of being blown by the wind. signage_catalogue.jpg ETA 6 months
Door or gate Polygon Door, if visible from imagery. door_catalogue.jpg ETA 6 months
Roof shape Classification Flat, low/moderate/steep pitch.   ETA 6 months
Fire truck Detection Useful for calculating distance to nearby fire station.   ETA 6-12 months
Fire hydrant Detection Useful for calculating distance to nearby fire hydrant.   ETA 6-12 months
Roof material Classification E.g. metal, shingle, tile.   ETA 6-12 months
Roof age Regression Time since roof was remodeled the last time.   ETA 6-12 months

3.3 Customization and other items

  • Improving accuracy of existing classes. We usually can reach 90%+ accuracy of direct work on a specific class in 2-8 weeks given priorities specified by the client.
  • New object types. We can add also add new object classes upon request. It usually takes between 1 and 3 months to support a new object class.

4 How Tensorflight can deliver data to me?

  • Web dashboard: Web dashboard that allows users to request new processing or view and edit results of our analyses. Please create an account and explore the dashboard at https://tensorflight.com/app .
  • GeoJSON: Standard format for describing GIS polygons easily integrated with web mapping tools available via the API accessible via an API.
  • Shapefile: Shapefile is one of standard GIS formats that can be easily imported into tools like ArcGIS or QGIS available via the API of the dashboard.
  • Vector tiles: URLs to slippy map tiles with vectors that can be overlaid on top of web based mapping tools available via the API. Example tile.
  • PDF export: Detailed PDF report per single area. Example pdf.

4.1 Output type

4.1.1 Bounding box

Rectangle around the object - min/max latitude/longitude.

Using this output type you can estimate object dimensions - e.g. bigger HVAC or bigger trees pose higher risk in case of a heavy wind. Based on object location we can also understand magnitude of the risk - e.g. tree or wind-borne debris dangerously close to the property is more likely to cause damage.

  1. Example

    tree_catalogue.jpg

4.1.2 Polygon

List of latitude/longitude points wrapping around the object.

This allows us to compute detailed square footage of the object. For example, using this data we can compute square footage of the property or percentage of the roof covered by solar panels or glass.

  1. Example

    degradation_catalogue.jpg

5 How to tell TensorFlight what do I want to analyze?

  • Email: If you are still evaluating our service you can simply send an email to kozikow@tensorflight.com with a list of properties you are interested in.
  • Send address via API: Users may send us an address they are interested in via the API.
  • Send location via API: Users may send us a boundary polygon of an area they want analyzed as a list of longitude/latitude points - e.g. Parcel boundary.
  • Dashboard address input: Users may input an address via the TensorFlight dashboard.
  • GeoTIFF upload: If you would rather to capture your own imagery you may also send us GEOtiffs. We support upload via the dashboard, API or fetching from services like Amazon S3.
  • DroneDeploy app: You can easily send us Drone data for processing via the DroneDeploy app.

6 Geographical coverage and Freshness

Freshness and geographical coverage of results offered by TensorFlight ultimately depends on the imagery we are analyzing.

For areas like metro areas in USA we usually can automatically find high quality imagery less than a year old via our network of imagery partners. For remote areas or other countries, we may need to compromise on accuracy or image freshness.

For detailed coverage and quality refer to the following table:

Location How detailed is the base imagery? How fresh is the data about the property?
Metro, USA/AU High - Aerial 2-3 months
Remote, USA High - Aerial 6-12 months
Metro, first world countries High - Aerial 2-12 months
Metro, others Low - Satellite 2-12 months
Remote, others Low - Satellite Up to 3 years

7 Latency

If we previously analyzed the property we can return full results in less than a second. Our long term goal is to have a database of all global properties.

If we never seen the property, the full analysis may take between seconds and multiple minutes depending on various factors.

8 Pricing structure

  Silver plan Gold plan Platinum plan
Properties analyzed per year 0-250K 250K - 1M Unlimited
Support Best effort PST working hours 24x7
Continuous property monitoring No No Yes
Price $2500 per 1000 properties $1M subscription per year $5M subscription per year