LONDON, SEPT. 17 – Mindtech Global Ltd, a UK based start-up, has announced the availability of Mindtech Chameleon Simulator, creating synthetic vision datasets for training neural networks, and Mindtech Chameleon AI Tools, providing end to end data management for deep learning systems.
Effective training of neural networks for visual processing requires very large datasets i.e. images with ground-truth annotations; access to these datasets is a major barrier to entry for most companies. The Chameleon Simulator economically creates unlimited, unbiased training data. A wide range of fully accurate annotations, including pixel perfect masks, precise range data, and derivatives such as velocity are easily generated.
Real-world data has many limitations which are overcome by Chameleon’s synthetic data: modelling of difficult edge cases, accurate synthesis of customer’s target system (lens, sensor, processing distortions) and enabling datasets free of privacy/GDPR issues.
Market and application optimized
Optional market-centric packs have been created to allow customers to rapidly create environments suitable for automotive, unsupervised machines, retail and security scenarios. Custom packs are created on demand.
Chameleon AI Tools simplify data wrangling tasks. They manage the merging, verification and augmentation of datasets for use in industry standard frameworks such as TensorFlow and Caffe2. The tools report and visualize relevant statistics for results analysis.
The use of synthetic data improves accuracy of neural networks, can actively reduce bias and vastly reduce the amount of “real” data required, saving time and money.
Says Chris Longstaff, VP Product Management, Mindtech: “Chameleon Tools enable everyone to bring innovative solutions to market. The ability to reduce bias is an important part of our company’s vision to allow for the ethical use of AI.”
“Sufficient quantities of diverse, high-quality, labelled images are critical for training and validating today’s visual AI solutions,” said Jeff Bier, founder of the Embedded Vision Alliance. “Synthetic images, such as those created by Mindtech’s Chameleon toolset, can ease the challenge of sourcing large quantities of labelled real-world images. This approach is especially interesting for applications in which developers require images that are difficult to capture in the physical world – for example because the images would be expensive to stage.”