Security - How HGH brings trust and performance into its AI video monitoring solutions
Impact
Challenges
Scaling complex data labeling
Control labels quality efficiently
Protect sensitive data
Solutions
A powerful tool able to label low-quality data
Robust quality review workflow
Comprehensive project management & tracking
Context
Since 1982, HGH Infrared Systems designs, develops, assembles, and sells electro-optics systems for industrial, civil, defense, and security applications. HGH customers use their products for monitoring, industrial thermography, and test and measurement applications. Today, the company has established itself as an international brand in terms of innovation in infrared technology, through the development of multiple advanced sensors.
Their catalog of 300+ creators in various categories from gamin to sports and beauty generates +10B monthly views across all social platforms, including Youtube, Facebook, Snapchat and Instagram.
How it all started
The Machine Learning Genesis
For years, HGH has been closely monitoring technological developments in the field of AI, so it could bring it into its line of products when ready. In 2019, after their successful participation in a European project aiming to use AI to help with drone surveillance, HGH decided to start integrating AI within their own systems. Three years later, HGH is using advanced AI models to provide its customers with more accurate, reliable, and efficient video monitoring capabilities.
One of the main pains of the organizations using video recognition for video monitoring is the number of false positives. This is cause for frustration and lost time, and therefore it is a key differentiator in the industry between us and our competitors, about who can provide the most reliable solution. They needed more reliability with less false positives.
Until 2019, expert systems were the best solution we had. The processing power needed to use neural networks to tackle our use cases was not accessible to us yet. But then, the development of a new generation of GPUs finally opened the doors, and we knew that AI could bring some new level of performance that would greatly benefit our customers.
Challenge
Scaling complex data labeling
Through 2019 and 2020, HGH managed its program in-house using a tool they had developed. The early results are encouraging, but they quickly realized that they will need to annotate a lot more data in order to get the performance they are after.
They were using interns during the school holidays to do the annotation work, using a tool they had developed. If it was a good start, it quickly showed limitations. The work on annotating data would only move forward during the school holidays, and the quality of labeling would not be optimal.
The videos and images we need to label can be very complex to interpret. Often, it requires some experience to figure out what it is that we are looking at. Therefore, I have to spend a lot of time reviewing the labels and helping our interns improve their labeling work. The problem is that we also need to accelerate and always label more data. So I knew that we needed to find a better way to do it.
:format(webp))
:format(webp))
An example of the type of images that HGH is working with. In this case, the shape of the object indicates that this is a sailing ship.
:format(webp))
In this example, only a light bump is visible. It’s only by looking at several frames of the video that the labeler can figure out that this is a moving object of dignificant size, a sailing ship, and not a wave or a bird.
In order to scale the volume of annotated data and improve labeling quality, HGH started looking for a professional data annotation solution.
Solution
Data security and IP, a critical imperative in the security industry
Datasets used for training ML models are a very valuable commodity in the security industry. There is a market for it, and many competitors would not hesitate to buy stolen datasets to run their R&D at a lesser cost. Data security was a key requirement for HGH when looking at vendors.
For security, project, and data governance, Kili offers three types of deployment to its customers in order to meet their requirements:
Regarding quality, Kili’s consensus feature enabled us to easily align annotators on our labeling guidelines, get them to cooperate on the most difficult cases, and ramp up their work quality much faster. This was one of the key differentiators that led us to work with Kili.
Data security is critical to HGH. We are dealing with all types of data, some that are classified. We also need to protect our R&D as this is what puts us ahead of the competition. Kili was one of the few actors able to provide us with the level of security and data governance we needed.
KILI WAS IDEALLY POSITIONED TO MEET HGH CRITERIA
Full SaaS solution, Hybrid deployment: the data and Kili’s platform are both hosted in customers’ servers, 100% on-premise.
Impact
Getting ahead of the competition and increasing customer satisfaction
When it comes to monitoring, It is very important to detect everything, while having a very low number of false alarms so as not to overload the operators, who more and more have to monitor several systems at the same time, or have other complementary tasks. The exponential increase in the amount of annotated data has mechanically translated into an improvement in both detection performance and the rate of false alarms.
Bringing AI into their systems and line of products is not unique to HGH, but the company managed to get a head start. It means improving its competitive advantage in the market by launching more robust products than its competitors, and better meeting its customer's highest expectations.
Kili helped the HGH team by enabling them to scale exponentially the number of data they can annotate in a short period of time and to better control their labeling quality and review workflow, leading to better AI performance. HGH is going to continue focusing on data quality and model performance improvement while also bringing AI to more products.
We increased by 30% the number of annotated data in just 2 months' work, compared to the last two years. It accelerated dramatically the performance improvement of our models and therefore the strength of our monitoring solutions value propositions.
Related Resources
AI For Compliance: What, Why and How?
Kili’s guide to the new AI paradigm
Opinion Classification with Kili & HuggingFac...
Get Started
Get started! Build better data now.