Self-Supervised Scene Depth Estimation from the Wild
Project Description

Scene depth is a fundamental element for an AI system to perceive the environment and more accurately perform various scene understanding tasks, such as detection, segmentation, and SLAM. In this project, we aim to develop a deep learning system that is able to estimate scene depth from videos. Different from existing works requiring expensive depth data as supervision to train the system, the proposed project is totally 3D ground-truth-data free. The whole system can be end-to-end optimized in a self-supervised manner.

Supervisor
XU Dan
Quota
2
Course type
UROP1000
UROP1100
Applicant's Roles

The applicants will be responsible for several parts:
(1) reading related papers in the literature and understanding the context of the proposed task
(2) developing a deep-learning-based self-supervised depth estimation system under the supervision of the supervisor
(3) writing a research report or a paper based on research experiments conducted and results achieved

Applicant's Learning Objectives

The learning objectives of the applicants include:
(i) having a deeper understanding of the task of self-supervised scene depth estimation
(ii) managing to use deep learning platforms and implement the proposed ideas for the task
(iii) developing research skills in computer vision and deep learning

Complexity of the project
Moderate