Black boxes, unexpected output and general failures are all pitfalls we had to watch out for while building a real-time computer vision pipeline. In this presentation we show you how we setup our real-time pipeline with focus on performance. Specifically, we talk about the implementation and use of our logging module.
We have been developing an automated camera solution, in which we use video input, neural networks for player detection, sport-specific decision making and a PTZ-camera to deliver an automated livestreaming application. The complete pipeline has to work in real-time with a maximum delays of 300ms. Within the 300ms we have to fetch the video streams, run several neural networks, apply business rules and control the PTZ. Over the past year, we have had to refactor code, remove complete modules or change python packages due to delay constraints. During the talk we will show the insights we gained during the process. Specifically, we will go into: