Jonathan Sauder: Unsupervised 3D Mapping from Video

 As deep learning has revolutionized many areas of computer vision in recent years, conventional computer vision methods have held up in 3D mapping tasks: for many applications, classical photogrammetry yields more accurate results than methods involving deep learning. However, in challenging environments with challenging lighting conditions and dynamic objects, many of the assumptions underlying conventional methods are violated. In such environments, deep learning can produce great results when classical 3D mapping completely fails.

This talk introduces unsupervised deep learning-based 3D mapping, the challenges associated with it, and ways to overcome some of them. In particular, a link with unrolled optimization in inverse problems is established, framing the learning-based 3D reconstruction process as an iterative procedure that can compete with the precision of conventional 3D mapping. We demonstrate the effectiveness of using learning-based 3D mapping for coral reef monitoring, delivering unprecedented scale and robustness.

Jonathan Sauder is currently a PhD Student in Deep Learning at the Environmental Computational Science and Earth Observation (ECEO) Laboratory Ecology at the École Polytechnique Fédérale de Lausanne (EPFL) in Switzerland. He completed his Bachelor’s degree in computer science at the University of Potsdam, and his Master’s degree in computer science with a focus on machine learning. In particular, his areas of interest were the intersection of optimization theory and signal processing with neural networks, as well as large-scale 3D deep learning.