Self-supervised Monocular Depth Estimation with Structure Regularities

We offer a master (or guided research) topic in the field of autonomous driving, which aims at providing accurate dense depth maps in a self-supervised manner.

Abstract

Self-supervised networks have achieved impressive performance in monocular depth prediction, but those depth maps cannot capture some small but important information like road-bump. To address this issue, we will propose a new framework generating more useful information rather than a rough depth map for the autonomous driving task.

Resources and Literature

• [1]. Godard, Clément. et al. "Digging into self-supervised monocular depth estimation." ICCV2019.
• [2]. Guizilini, Vitor. et al. "3D Packing for Self-Supervised Monocular Depth Estimation." CVPR2020
• [3]. Ranftl, René. et al. "Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer." TPAMI2020.
• [4]. Guizilini, Vitor. et al. "Semantically-Guided Representation Learning for Self-Supervised Monocular Depth. ICLR2020
• [5]. Zhao, Wang, et al. "Towards Better Generalization: Joint Depth-Pose Learning without PoseNet^?." CVPR2020
• [6]. Irani, Michal, et al. "Parallax geometry of pairs of points for 3d scene analysis." ECCV1996.

Requirements

• A basic background in Multi-View Geometry (SLAM/SfM).
• Familiar with self-supervised depth prediction projects.
• Experience in training networks.

If you are interested in this topic, please contact us via email.

Yanyan Li

Federico Tombari