Abstract: This talk will have two parts. First, I'll discuss what we've learned from Argoverse competitions in 2020 and 2019. We'll analyze the strategies used by the top scoring teams in 3D tracking and Motion forecasting, and examine situations where there is still room for improvement.
In the second part, I'll discuss the "inflation" of 2D instance segmentations into 3D cuboids suitable for training 3D object detectors. With the help of an HD map, 2D instance masks can be converted into surprisingly accurate 3D training data for LiDAR-based detectors. We show that we can mine 3D cuboids from unlabeled self-driving logs and train a 3D detector that outperforms a human-supervised baseline.
Bio: James Hays is an associate professor of computing at Georgia Institute of Technology since fall 2015. Since 2017, He also works with Argo AI to create self-driving cars. Previously, he was the Manning assistant professor of computer science at Brown University. He received his Ph.D. from Carnegie Mellon University and was a postdoc at Massachusetts Institute of Technology. His research interests span computer vision, robotics, and machine learning. His research often involves exploiting non-traditional data sources (e.g. internet imagery, crowdsourced annotations, thermal imagery, human sketches, autonomous vehicle sensor data) to explore new research problems (e.g. global geolocalization, sketch to real, hand-object contact prediction).