Zoom: https://illinois.zoom.us/j/85139410235?pwd=Mmk2ME1vZjFQRjVBM3JLUXgwb0Z1dz09
Abstract:
Autonomous driving technology has advanced greatly in recent years with the application of machine learning (ML) methods. However, treating it as a general and abstract ML practice will inevitably share the typical drawbacks in ML, such as inferior performance in tail cases, data & label hungry, etc. But the good news is that we might have posed the problem a bit harder than we should have.
In this talk, I will introduce a previously ignored, simple yet highly practical viewpoint of the data in the autonomous driving application: repeated unlabeled traversals of the same scenes. It exploits the nature of driving: most of the driving scenarios are not completely new in that they are traversed by cars multiple times, and the unlabeled sensor data from them can be obtained almost free.
This talk will mainly present our results from two different angles from this viewpoint: 1) Hindsight (ICLR22): How to use unlabeled multiple traversals even in test time to significantly enhance 3D perception, especially in tail cases; 2) MODEST (CVPR22): Is it possible to train a dynamic object detector without a single human label that is competitive with a supervised approach? With these preliminary explorations, we will discuss the future directions on further such viewpoints on other autonomous applications and other ML tasks.
Bio:
Yurong You is a final-year Ph.D. candidate in Computer Science at Cornell, advised by Prof. Kilian Q. Weinberger and Prof. Bharath Hariharan. He received his bachelor's degree from Shanghai Jiao Tong University (ACM honors class) in 2018. His research is mainly focused on 3D computer vision and machine learning, with an emphasis on learning with limited labels and applications in autonomous systems. Some of his highlight works include Pseudo-LiDAR++ (ICLR2020), Hindsight (ICLR2022) and MODEST (CVPR2022).
Part of the Illinois Computer Science Speakers Series. Faculty Host: Shenlong Wang
Join us on Zoom (meeting ID:851 3941 0235 , password: csillinois).
