CVPR Daily - Wednesday

“ When we use an autonomous driving stack in the real world, the simplest thing we can do is run detection and tracking, which will give the past trajectories of the objects around the robot we’re trying to control, ” Xinshuo explains. “ Currently, although they are deemed good enough, detection and tracking still give some errors. That’s why this work is important. ” A great deal of engineering effort is involved when building a robust autonomous driving pipeline , including detection and tracking and trajectory prediction modules. It is not just about solving a single task but building a whole system across a wide range of tasks . “ There are many state-of-the-art machine learning models out there, ” Xinshuo points out. “ That’s been helpful because I’m not developing detection and tracking from scratch; I’m leveraging all these open-source methods and algorithms already available. It would have been very challenging without that. Although it requires much engineering in the upstream perception modules, our main focus is on the prediction side and the challenges there. ” This work is an essential step toward integrating all kinds of autonomy stacks . There are many research papers on autonomous driving domains such as detection and tracking and trajectory prediction. However, Xinshuo says a more robust stack that has been integrated seamlessly across the autonomous driving stack is needed to build a unified framework that can solve the entire problem in one go . This work is the first part of that, and the next part is building something based on it to achieve a more robust autonomy stack as a whole. One crucial aspect of this problem that has been underexplored so far is how 4 DAILY CVPR Wednesday Poster Presentation