The talk given by Stefan Baur in KUIS AI Talks on November 12, 2024.

Title: Self-Supervised 3D Scene Flow for LiDAR Object Detection Without Human Annotations

Abstract: Human annotations for lidar data are expensive and do not scale well. At the same time, temporal consistency and motion are cheap and strong supervision signal present in sequences of lidar point clouds. In this talk, I will first describe how to obtain accurate self-supervised 3D scene flow from lidar point cloud sequences.
Then, using the motion cues from the scene flow, I will demonstrate how we are able to train self-supervised lidar object detectors that generalize from moving to movable objects, using only temporal consistency (tracking) as supervision, but no human annotations.

Short Bio: I joined Mercedes-Benz R&D and University Tübingen Autonomous Vision Group (Andreas Geiger) in 2018 as a PhD student. First, I worked on realistic lidar simulation for autonomous driving. Later, I pivoted towards motion-based self-supervised learning for lidar point clouds. In 2022, I became a full-time ML engineer at Mercedes-Benz. At the moment, I am working on online HD map learning.

source