Abstract

For monocular depth estimation, acquiring ground truths for real data is not easy, and thus domain adaptation methods are commonly adopted using the supervised synthetic data. However, this may still incur a large domain gap due to the lack of supervision from the real data. To this end, we develop a domain adaptation framework via generating reliable pseudo ground truths of depth from real data to provide direct supervisions. Specifically, we propose two mechanisms for pseudo-labeling: 1) 2D-based pseudo-labels via measuring the consistency of depth predictions when images are with the same content but different styles; 2) 3D-aware pseudo-labels via a point cloud completion network that learns to complete the depth values in the 3D space, thus providing more structural information in a scene to refine and generate more reliable pseudo-labels. In experiments, we show that our pseudo-labeling methods improve depth estimation in various settings, including the usage of stereo pairs during training. Furthermore, the proposed method performs favorably against several state-of-the-art unsupervised domain adaptation approaches in real-world datasets..

Bio.

Wei-Chen Chiu received the B.S. degree in Electrical Engineering and Computer Science and the M.S. degree in Computer Science from National Chiao Tung University (Hsinchu, Taiwan) in 2008 and 2009 respectively. He further received Doctor of Engineering Science (Dr.-Ing.) from Max Planck Institute for Informatics (Saarbrücken, Germany) in 2016. He joined Department of Computer Science, National Chiao Tung University as an Assistant Professor from August 2017 and got promoted to Associate Professor in July 2020. From May 2021, he was also hired as a Joint Appointment Research Fellow by the Mechanical and Mechatronics Systems Lab of  Industrial Technology Research Institute, Taiwan. His current research interests generally include computer vision, machine learning, and deep learning, with special focuses on generative models, multi-modal perception, and 3D recognition.

All faculty and students are welcome to join.