Speaker:

Dr. Renaud Marlet, Senior Researcher, École des Ponts ParisTech

Abstract

Surface reconstruction from point clouds has made huge progress in the last years, highlighting the potential of neural network for this task. In this talk, I will present a relatively simple but efficient supervised method to address this problem (POCO [CVPR 2022]). I will also show how this reconstruction task can be learned without supervision, using only sensor location information [ICPR 2022], and how it can be used to construct rich point features in a self-supervised manner, towards both semantic segmentation and object detection (ALSO [CVPR 2023]). Last, I will present how this self-supervised reconstruction can also be used as a signal to perform domain adaptation for the semantic segmentation of lidar point clouds (SALUDA [3DV 2024]).

Biography

Renaud Marlet obtained a PhD in Computer Science in 1994 from the University of Nice-Sophia-Antipolis (France). He has held positions in both the academic world (researcher at INRIA) and the software industry (project manager at Simulog, deputy CTO at Trusted Logic), where he has successively worked on programming languages, software engineering, security and computational linguistics. Since 2009, he has been a Senior Researcher at École des Ponts ParisTech (ENPC) in the Computer Science Laboratory (LIGM), leading the IMAGINE research group (2010-2019) and working on computer vision, using both 2D and 3D data, towards robotic applications in the context of civil engineering. Since 2019, he has also been Principal Scientist at valeo.ai, working on safe driving and autonomous vehicles.