We are thrilled to announce our investment in Cartica AI. Cartica AI brings a new approach to computer vision in automotive applications, leveraging proprietary technology developed over more than 10 years.
This is the first round of financing for the newly created company, which is a spin-off from the company Cortica that has been developing computer vision based algorithms and software since 2007. While the core technology can be applied to many different use cases, Cartica AI is solely focused on the automotive industry.
A vehicle computer vision system consists mainly of three components:
-
Sensor (e.g. camera)
-
Chip
-
Software (e.g. for object detection and classification)
Cartica AI has developed software that enables vehicles to identify objects in their environment and make informed decisions. The company’s first industrialized product – “Cartex 1.0” – is a perception stack for forward facing ADAS (Advanced Driver-Assistance Systems) cameras. It provides a cost-effective alternative to the existing systems based on deep learning, and currently supports a range of applications which will play a significant role in NCAP ratings from 2020 onwards, including Automated Emergency Braking, Adaptive Cruise Control or Traffic Jam Assist.
Cartica AI introduces a breakthrough technology – the algorithm takes a fundamentally different approach in comparison to deep learning by mimicking the way in which the brain works. The same way a brain emits a neural reaction when exposed to a combination of stimuli, Cartica’s system translates visual information from cameras, LIDAR, radar and audio into highly compressed text representations, called concept signatures. Concepts within the automotive context can for example be object classes, such as pedestrians, stop signs, bicycles, lane markings, trucks etc. This process makes complex information and large amounts of data easily searchable, scalable and better organized for deeper insights.
Cartica AI introduces a series of important technological developments for an OEM:
-
Low computing power: Cartex 1.0 requires only 0.5W of computing power to operate on standard chips.
-
Contextual understanding: the system understands both the concepts and their context, giving it the power to predict how the surrounding objects will behave. For instance, after being fed with random traffic footage, the system can learn on its own to predict that a bus, next to a bus stop full of people waiting to get into the bus, is 99% likely to remain in that position for the next seconds. In comparison, deep learning algorithms focus on recognizing objects by their shape without context.
-
Unsupervised self-learning: the system can operate and learn in real time on its own, without requiring a tagged training set. When the system is exposed to new situations, it learns organically by comparing them to the stored signatures and finding similarities. Cartica’s AI technology creates concepts with high accuracy in a multitude of weather and lighting conditions, covering programmatically the long tail of edge cases. Deep learning systems, on the contrary, require pre-training with massive data through very manual processes and struggle under new and unfavorable conditions.
-
Predictability: the flat, hierarchical architecture of the system makes it transparent – the outcomes can be traced to their root causes and explained. This allows to easily identify and fix issues without compromising all the data collected previously. Thus, the customer can generate, improve and test any concept without the need to retrain the network. On the other side, deep learning systems are often compared to a black box and the outcomes are hardly explainable.
-
Second source: Given Cartica’s fundamentally different approach to object classification it may be used in parallel to other deep learning solutions to validate the classification and create redundancy.
-
Collaboration: Cartica’s technology is able to cooperate with other AI agents and modules, enabling knowledge sharing between vehicles and infrastructure (V2X) in an efficient way thanks to the lightweight representations.
We believe Cartica AI has the potential to become the category leader for the next generation of automotive computer vision algorithms. Cartica’s approach has a series of significant technological advantages over deep learning techniques as described above. Although it seems that the market is already populated, Cartica is the only company that is industrializing signature-based technology for ADAS systems. Our conviction in its future success is reinforced by the following:
-
By addressing mandatory NCAP requirements with a low cost, mass market solution, Cartica could offer a superior solution on the market.
-
The necessity for OEMs to validate autonomous systems’ decision making and create redundancies between sensors opens a great opportunity for Cartica to be the “second source” of truth in the perception stack. Thus, the product can be used by OEMs in parallel to other deep learning based systems following a differentiated approach.
-
Cartica AI’s ability to quickly train and retrain new concepts (objects) enables an incredibly short time to market. The need to retrain a complete deep neural net struggling to validate shortly before starting to produce a vehicle is a huge pain point.
These were the main reasons for us to invest in a company that holds great potential to become a success story. We believe Cartica AI will thus play a key role in the development of Autonomous Driving and Advanced Driver-Assistance Systems (ADAS) by OEMs, allowing them to design safe and predictable vehicles that fulfill the strict requirements of the market. We are looking forward to supporting Igal and the rest of the Cartica AI team in further growing the business.