DeepScale – a San Francisco-based specialist in ‘machine learning’ software – was purchased by Tesla in October 2019 for an undisclosed amount to assist in development of autonomous driving technology, according to a reported by Electrek.
DeepScale had already developed ‘Carver21’ – an artificial intelligence software for self-driving vehicles – prior to Tesla’s purchase, however it is unclear if that same software lays the foundations for Tesla’s new autonomous driving software.
Tesla's patent application – named “Systems and Methods for Training Machine Models with Augmented Data” – aims to improve the way Autopilot software uses its eight car-mounted cameras to identify objects in an environment – also known as ‘3D labelling’.
“In typical machine learning applications … sets of images used for training computer models may represent objects captured [in] many different capture environments [with] varying sensor characteristics,” the patent explains.
“These sensors may also differ with respect to different extrinsic parameters, such as the position and orientation of the imaging sensors (cameras) with respect to the environment as the image is captured. All of these different types of sensor characteristics can ... make it more difficult to properly train a computer model.”
Translated, this means an autonomous driving system may receive programming to recognise characteristics of a particular object, but these characteristics may not always match what cameras register in certain environments or situations, confusing the system.
The patent lists focal lengths, lens types, pre- or post-processing, different software environments and sensor array hardwares as possible causes of this discrepancy.
Tesla and DeepScale’s solution – according to the patent application – lies in introducing pre-augmented imaging of how objects appear in varying environments to the software. This would be done in hopes that the software may learn what to expect in various conditions and become able to make accurate corrections for differing environments itself, resulting in a system that can:
“[Generate] an augmented image for a set of augmented images by … modifying the image(s) with an image manipulation function that maintains camera properties of the image,” and "train the computer model to learn to predict the training output based on an image training set including the images and the set of augmented images.”
Above: an image from Tesla's patent
CarAdvice understands this to mean Tesla’s new software would be able to gather information of a vehicle’s environment, calculate how the conditions may affect the way an imaging sensor captures an object, augment the image captured and update the software’s parameters accordingly to recognise objects.
The software and updates to Tesla’s Autopilot and 3D labelling technology could spell a pivotal step in the move towards self-driving cars for the brand.
Tesla’s purchase of DeepScale comes as one of many made by the car making industry to snatch up autonomous driving tech companies, with Toyota investing in Silicon Valley-based Pony.ai in February and partnering with Chinese company Momenta – that develops high-definition mapping software crucial for self-driving – cars last month.