ISP adaptation for computer vision

About this session

In this presentation we focus on a practical case of using ISP to collect training data. Widely asked question in computer vision society is how to tune ISP for training data collection and how the existing metrics for human vision extrapolate to effectiveness of object detection. The other concern in the autonomous driving community is how to deal with the dynamic behaviour of camera and ISP.

Suggested approach addresses both of the matters in a data driven way. As the result of collaboration of Pony.AI and ON Semiconductor we propose an ISP based solution which does real-time hardware accelerated data augmentation while capturing images for training set. Having such an augmented dataset makes machine learning networks robust to different variations of how the same object can be captured by an automotive camera system.  The presentation will also briefly cover how ISP can be utilized for adapting wide angle camera solutions for navigating in spaces like parking lots.