Domain Adaptation in Computer Vision

CSE 291 A00, Winter 2020

Instructor: Manmohan Chandraker
Email: mkchandraker [AT] eng [DOT] ucsd [DOT] edu

Lectures: WF 5-6:20pm in EBU3B 2154
Instructor office hours: Thu 4-5pm at CSE 4122

TA: Yu-Ying Yeh (
TA office hours: Tue 2-3pm in EBU3B B215


Computer vision has made rapid progress in the era of deep learning. This is largely attributed to the availability of large-scale labeled data, coupled with GPU computation. Yet, computer vision models trained on a domain, say daytime images, often do not generalize to new domains, say images acquired at night. It is expensive to label data for all possible scenarios, but unlabeled data is easier to obtain. In this course, we will study concepts in unsupervised domain adaptation, with applications to various computer vision problems, such as image classification, semantic segmentation, object detection, face recognition and 3D reconstruction.


This is an advanced class, covering recent research progress. Graduate students with a strong interest in computer vision may enroll. Prior background in computer vision and machine learning is required, preferably through research experience or as covered by CSE 252 and similar offerings. Students are encouraged to contact the instructor if unsure about meeting any criteria for enrollment.

Course Format and Requirements

The course will consist of lectures by the instructor on concepts for domain adaptation and applications in computer vision, as well as presentations by students on an assigned paper. Students will be required to do a project on a focused topic. Wide flexibility is available for the choice of topic, but should be discussed with the instructor beforehand. The instructor can also provide a topic. Projects should be in groups of four. The best projects may consist of conceptual development or practical implementation that pushes the boundaries of the state-of-art.

Participation has two components: reviews and in-class discussions. The day before every class that covers papers, a brief review of one of the papers to be discussed must be submitted. Besides, students are expected to actively particpiate in discussions in the class.

Grades will be weighted as 40% for project, 20% for in-class presentation, 20% for final exam, 10% for reviews and 10% for participation. The goal of the course is to understand the current state of domain adaptation and gain appreciation of its limits and potential.


The course will cover a diverse range of topics in domain adaptation, including: These methods will be applied to several problems in computer vision, such as:


Jan 08: Introduction Jan 10: Overview Jan 15: Instance reweighting Jan 17: Maximum Mean Discrepancy Jan 22: MMD in deep networks Jan 24: t-SNE Jan 29: Entropy Jan 31: Landmark selection Feb 05: Correlation alignment Feb 07: Optimal transport Feb 14: Architecture Feb 21: Generative Adversarial Networks Mar 04: Domain Adversarial Learning Mar 06: H-divergence Mar 13: Open Set Adaptation


Manmohan Chandraker
Last modified: Fri, Feb 14, 2020