Looking For Information On Thirdeye

They are a new supervised machine learning computer vision company in the UK working on tracking human movement in enough detail to identify a thief before he/she steals as a first commercial opportunity.


We have not heard of them previously, but have reached out to see if we can setup a conversation to gather some information.

How did you hear of them? (notice you are also in the UK).

Response from Adrian:

Saw them at the Entrepreneur First demo day - it’s an accelerator in London.

As far as we can tell, no one in the market can identify specific, detailed human behaviour in real time from CCTV footage - e.g. someone taking an item off a supermarket shelf and placing it in a bag as opposed to a basket (ie. shop lifting). Do you know anyone who claims to be able to do this? BRS / Grey can detect anomalous behaviour but not something so specific.

I spoke with Thomas from Thirdeye today to get some more information.

For the first phase, Thirdeye is looking at the retail theft/loss prevention application. This is a good approach, it is a well defined (and measured) problem in the retail space.

The system is still in the early stages of development. They're using a machine-learning approach, taking videos that Loss Prevention operators have flagged as containing suspicious behavior, and feeding that into a neural learning network.

Right now the system needs about 30-40 pixels vertically on a person (6ppf) to be able to see actions it looks for like concealing an item, or grabbing something.

They're using advanced GPUs, and still working out the amount of processing resources needed for a camera, so hard to say how much hardware is required in a typical system yet.

It sounds like a commercialized version is still several months to a year away.

Sounds really difficult to achieve. I often do not use baskets and care everything in my hands or pretty often I see people put things into the strollers' baskets or sometimes people use their bags as a basket.

I agree that it sounds very difficult. One downside of neural-network learning from video clips is that you can never fully control, or understand, what the system is learning. You can have it flag things in sample videos, or in actual clips, but it can be very difficult in cases like this to have a crisp way to define "show me less of that, and more of this".

The upside for an application like this is that it doesn't have to be perfect, it just needs to assist the full-time loss prevention people by enough to justify the price. If the system can successfully spot some suspicious activity, without throwing a lot of false alarms, and act as a pre-filter for the LP team, it might be cost effective.

However, I think they will find that "the public" is a strange animal, and people do weird stuff with no bad intentions at all. Detecting things isn't all that hard, but intelligently ignoring similar items or actions is.

I know it first hand, we teach our artificial neural network on 500K plus people and constantly increase this dataset. But we deal with faces and it is easier to gather needed material.

ANN is also used for search of sample objects (i.e. find me this car or exactly similar on these videos) and it works but in their case, they want deal with something like this

in a complex environment with different angles, different light, people in winter clothes... I do not know, it will be really cool if they will be able to make it.

IMO behavioural analytics and biometrics are strange things. Sometimes it can show great results but so far, it is very unpredictable.

To play devil's advocate, there are some dozen companies doing retail LP video analytics. It's a fairly crowded space, and it takes a lot of work to integrate / get some in various retailers.

And it is not a space where there has been that much success. The big name is ShopperTrak who got bought by Tyco for $175 million but that was mostly because they have been around forever and had a lot of accounts.

What is the technology differentiator here that is going to break them through against the various smaller players that had struggled for years?

Were you able to use cameras to identify shoplifting BEFORE it occurs (the claim), what is the next action of the system? Scramble people to the location? Flash them with camera LEDs? Make an announcement over the PA? Shoot a net over the person? LOL.

I feel the claim of identifying "pre-incident" is, and will remain, hype.

For shoplifting, you would likely alert the loss prevention people, have them review video to verify, and then apprehend the person in the store, or when they try to leave.

Most of the time you should have a little time between when the person conceals the item, and when they go to leave, so you don't need an instant response.

I don't think they intend to prevent shoplifting with a pre-crime approach, but instead to identify common actions in real-time so a human operator can review and apprehend before the person has time to leave the store with the good. As long as the item isn't damaged and can still be sold, that would be marked as a "prevention".

In 10 years of managing the LP function for a major retailer (involved in thousands of apprehensions), I just don't see it being that simple in a real-world application.

First the human CCTV operator checks what's going on and confirms or denies suspicious behaviour and if confirmed alerts security to make their presence known. If the theft has not yet occurred, the presence of the security guard will ensure it doesn't take place. If it has happened, the thief has been recorded on CCTV making the theft and can be apprehended.

A key thing that Third Eye will have to demonstrate is the percentage of alerts that are accurate. This is a common structural issue for video analytics.

If Third Eye's alerts are 50% 'right' in picking out shoplifting, great, retailers will probably be thrilled but if it 10%, not so much.

Also, what portion of shoplifting events will they hit vs miss? If a store gets shoplifted 10 times a day to use a round number, will Third Eye get 1 or 2 or 5 on average?

And what needs to be done with cameras? Are more needed? Do they need to be at different angles than currently positioned?

Related: Retail Surveillance Operator Secrets Revealed has a lot of interesting operational info related to handling shoplifters.