Posted by alphaatlas 9:05 AM (CST)
Thursday November 08, 2018
GPUs are already being deployed to track people outside of Teslas and on the streets, but apparently, the algorithms those GPUs run are limited to relatively low resolutions of about 608x608 pixels. To keep up with improving camera technology, researchers have come up with a new algorithm that can handle higher resolution video. "YOLO v2" downscales the large video frames first, and "performs a fast object detection" that it uses to intelligently crop the full resolution frame and send the pieces to multiple GPUs. This yields much better results than just downscaling a feed before processing, and is faster than processing the full frame. Check out a video of the algorithm in action below:
"Our method reduced the time necessary to process high-resolution images by approximately 20 percent, compared to processing every part of the original image under high resolution," Ruzicka said. "The practical implication of this is that near real-time 4K video processing is feasible. Our method also requires a lower number of server workers to complete this task."