Training Artificial Intelligence Systems

By on

Computer vision systems are driving some incredible developments.

Convolutional neural network (CNN) algorithms are most commonly applied to analyzing visual imagery. That involves the “classification, detection, [and] segmentation of entities in the 2D space and analyzing temporal information in the 3D space, such as Action/Video,” says PR Krishan, VP and Global Head, Enterprise Intelligence Automation, Tata Consultancy Services.

An article in Forbes notes that CNN can be applied to, among other things:

  • Autonomous vehicles, where cameras and sensors acquire images from the environment so that objects or boundaries can be detected for driver safety.
  • Healthcare, where 90% of medical data is image data that can be critical to developing new medical diagnostic methods.
  • Manufacturing, where equipment is monitored by computer vision to enable intervention before there is a machine breakdown.

Startup LexSet is building its company on a visual AI solution for object recognition and visual search, specializing in synthetic data generation. The company, a spinout of Intellectual Ventures, just won the Startup Lab Pitch Competition at GS1 Connect 2019. GS1 is the standards organization that facilitates industry collaboration to help improve supply chain visibility and efficiency through its standards.

The intersection between what a computer vision company does and who GS1 is may not be immediately obvious. LexSet co-founder Les Karpas explains the connection:

“GS1 is very interested in product identity and we improve computers’ ability to see products, count them, track them, and know where they are in the supply chain.”

Francis Bitonti, who has done significant work in the area of generative design of 3D-printed products, is also a co-founder of LexSet.

Fixing AI’s Image Training Data Problems

As with any other AI application, the success of computer vision artificial intelligence outcomes for use cases such as supply chain lies with training data. But there aren’t always enough training images or enough diversity of images for computer vision algorithms to work as well as hoped. For example, photographic training data may be missing different camera angles or suffer from poor lighting conditions that could affect end results.

Karpas believes that AI usage is rising at an exponential rate:

“But a lot of companies find that as they start to create AI solutions for problems, they are running into training data problems right away.”

That’s where synthetic data generation comes in, generating on-demand photo realistic synthetic data customized to any use case for the development of vision models from the data they create.

LexSet can take an object or a set of them from 3D CAD files and 3D scans and create thousands of images from all possible lighting conditions, camera angles and so on for the highest-quality training set, Karpas says.

“We can completely replace photos in a training set. And synthetically-trained models have a 15% accuracy increase over photo-based models because of the large volumes of data,” Karpas says.

Use Cases for LexSet

Take a supply chain example. A company with fast food chains may be interested in doing realtime inventory using the security cameras that are already in its restaurants. They want a solution that will work with those cameras to monitor the relative quantity of any given product on the floor at a given time, based on the identification of those products made possible by the use of synthetically generated training data. With that in place, they can know the right moment to send a resupply truck and to run the whole operation without having to buy additional hardware, Karpas says. 

“Inventory management is a big problem across the supply chain,” says Melanie Nuce, GS1 US innovation expert. Autonomous retail use cases can also be built around using the technology for robotics backup fulfilment and unmanned kiosks. “CVS is heavily deployed in automatic retail use cases, and that’s of great interest to GS1 members,” she says.

Other Uses of AI Training and Search Visualization

Another application that LexSet is piloting is related to a large contract furniture maker that uses its Similar Product Search technology.

With LexSet technology, the company now has a tool to sit on top of their catalog portal in competitive bid processes. The tool makes it fast and easy to identify products from its own catalog that are equivalent to products that other brands are submitting in their own bids. Recommendations are driven by the product type and category and visual similarities. “It would take a design team a week to do that manually,” Karpas says. This saves multiple man-hours in the process.

Once LexSet has a large quantity of 3D models in a particular industry, developing new 3D content is not necessary:

“We could provide value since we have reached a critical threshold in that base,” Karpas says. “Once we have tackled a vertical, it’s very rinse-and-repeat for other players in that vertical.”

LexSet’s technology also has the potential to replace the use of more expensive RFID chips in some scenarios, Karpas says.

Field tech workers for a customer in the industrial tools space, for example, were having a difficult time telling one tool apart from another, since they were very similar looking. The company tried embedding RFID chips in the parts, but the chips were getting destroyed by their usage in heavy field work applications – and they were costly, too. LexSet built a vision model to replace the chips for sustainable use and cost savings.

GS1 Partnership Possibilities

“We are working on how standards and emerging technology can be brought together to solve business problems in a new way,” says Nuce. “There’s a really good opportunity for GS1 to pair with LexSet as they expand their use cases.”

Simplista and Locai took second and third place honors in the StartUp Labs competition. Simplista tackles the product intake listing and maintenance process. Locai focuses on the eGrocery market by using automated robotics solutions to support fulfillment formats from in-store to dark store.   

Image used under license from

Leave a Reply

We use technologies such as cookies to understand how you use our site and to provide a better user experience. This includes personalizing content, using analytics and improving site operations. We may share your information about your use of our site with third parties in accordance with our Privacy Policy. You can change your cookie settings as described here at any time, but parts of our site may not function correctly without them. By continuing to use our site, you agree that we can save cookies on your device, unless you have disabled cookies.
I Accept