Advancing 3D Object Detection for Autonomous Driving and Aerial Systems

3D object detection in LiDAR point clouds is a fundamental capability for autonomous and aerial systems, enabling robust perception in complex environments. In this seminar, I will present a series of contributions that tackle key challenges in this field, including data scarcity, computational inefficiency, and domain generalization. I will introduce GraVoS, a gradient-based voxel selection method that improves training efficiency and detection accuracy; PatchContrast, a self-supervised pre-training framework that reduces reliance on labeled data; and WiSAR3D, the first large-scale aerial LiDAR dataset for wilderness search and rescue. Finally, I will discuss SFMNet, a novel sparse detector designed to capture both short- and long-range context in large-scale outdoor scenes. Together, these works provide new insights and tools for building efficient, generalizable, and high-performance 3D object detection systems for real-world applications in autonomous driving, robotics, and aerial monitoring.

Ph.D. student Under the supervision of Prof. Ayellet Tal.