Startup Image Intelligence 10X Accuracy Guarantee

An Australian startup, Image Intelligence, is targeting the biggest problem of video analytics, poor accuracy, with quite a claim:

Is this legitimate or a stunt? What does Image Intelligence aim to do? We spoke with Image Intelligence's founder Jimmy Lee [link no longer available] to better understand their offering and the accuracy guarantee.

Company Background

Image Intelligence is developing this internally, with 7 of the 8 employees on LinkedIn being software engineers or developers. 

The Image Intelligence team was originally part of Cammy. a cloud-based residential camera recording system, spun off in 2016 to focus on the development of the people detection analytics they developed. At Cammy, they were addressing problems with false motion-based detection alerts which lead to high video storage usage. This was causing higher-than-estimated storage cost for Cammy. To address the problem, they developed machine learning based analytics to identify people.

Only People / Pet / Vehicle Detection

Image Intelligence offers people, pet and vehicle detection, including face search. The video below shows a confidence score as people or vehicles move through the scene: 

They do not offer more conventional analytics such as tripwire, perimeter violation, people counting, etc.

Person Matching

Their "Match" programming attempt to find a person within the video based on providing 3 images of a person.

Image Intelligence also claims:

A match prediction is not only based on facial recognition, as often clear face shots are not available in security camera footage. Other features like age, gender and skin tone are used in this assessment.

No Recorded Video / No End User UI

Image Intelligence does not record video nor does it offer an end user UI. Image Intelligence receives video, detects relevant people, pets or vehicles, cataloges them and makes it available via API to access those results. An API [link no longer available] is available to integrate a video system by sending motion detection generated images or video to their cloud service, and receive a confidence score and "bounding box" data back from the cloud service.

Offline Version Available

An on-site version of the software is available through a SDK Program to run on GPU enabled machines, which needs to communicate with their cloud services. Pricing was not available for the offline version.

Pricing Per Frame, Very High

Image Intelligence charges per frame analyzed ($0.10 USD per 1000 frames), unlike most systems which charge per camera. This is a common model in image data-mining verticals and much less expensive than Amazon or Google (Amazon Rekognition, $1 per 1000 frames), Google Cloud Vision at $1.50 per 1000 frames).

However, the cost would be $3.60 for an hour of video at 10fps, which would be incredibly expensive. Even at 1fps, it would be $0.36 per video hour, or over $250 per camera per month.

3rd Party Integration

Image Intelligence says they have an integration with an older version of Network Optix, but did not clarify which version. Additionally, they are working on other VMS integrations.

Guarantee Examined

Every time you identify a mistake in a frame, they effectively credit you a penny ($0.01) for that frame (i.e., 10x what you paid for that frame - 1/10th of a cent). This is not worth enough to go back and manually review all of your video. If you are having serious issues with accuracy, like any other analytic, the best response is to stop using it as the guarantee does not 

Outlook For Company

While Image Intelligence has eye-catching marketing, their offering suffers from a number of problems:

• High pricing
• Limited analytics (tripwire, perimeter violation, people counting, etc.)
• Lack of video client web interface
• Lack of 3rd party VMS Integrations

To become a serious contender in video surveillance, they will need fundamentally reexamine how they approach the market.