Dave Tokic, Vice President of Marketing and Strategic Partnerships at Algolux, demonstrates the company's latest embedded vision technologies and products at the February 2017 Embedded Vision Alliance Member Meeting. Specifically, Tokic demonstrates the company's novel machine learning approach to addressing the costly and manual intensive challenge of tuning cameras for proper image quality (IQ) for both visual perception and computer vision, and for camera-based applications such as ADAS, autonomous vehicles, inspection, surveillance, etc. The demonstration shows CRISP-ML optimizing image quality on a Movidius (Intel) Myriad 2 reference platform with an OmniVision 2 Mpixel image sensor.
Today's IQ tuning process requires deep expertise in imaging and enough time and budget to manually tweak hundreds to thousands of parameters of a camera's Image Signal Processor (ISP) block to achieve the right settings. This can take weeks to months of time and must be done for each lens/sensor/ISP configuration. CRISP-ML applies a hyper-parameter optimization approach against target metrics (KPIs) for the image quality or computer vision accuracy you are trying to achieve, reducing the time down to hours or days and allowing you to test many different lens/sensor configurations to optimize system cost and performance.
