Semi-Supervised Domain Adaptation with Instance Constraints

TitleSemi-Supervised Domain Adaptation with Instance Constraints
Publication TypeConference Paper
Year of Publication2013
AuthorsDonahue, J., Hoffman J., Rodner E., Saenko K., & Darrell T.
Other Numbers3418

Most successful object classification and detection methodsrely on classifiers trained on large labeled datasets.However, for domains where labels are limited, simply borrowinglabeled data from existing datasets can hurt performance,a phenomenon known as “dataset bias.” Wepropose a general framework for adapting classifiers from“borrowed” data to the target domain using a combinationof available labeled and unlabeled examples. Specifically,we show that imposing smoothness constraints on the classifierscores over the unlabeled data can lead to improvedadaptation results. Such constraints are often available inthe form of instance correspondences, e.g. when the sameobject or individual is observed simultaneously from multipleviews, or tracked between video frames. In these cases,the object labels are unknown but can be constrained tobe the same or similar. We propose techniques that buildon existing domain adaptation methods by explicitly modelingthese relationships, and demonstrate empirically thatthey improve recognition accuracy in two scenarios, multicategoryimage classification and object detection in video.


This work was partially funded by the Deutscher Akademischer Austausch Dienst (DAAD) through a postdoctoral fellowship.

Bibliographic Notes

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Portland, Oregon

Abbreviated Authors

J. Donahue, J. Hoffman, E. Rodner, K. Saenko, and T. Darrell

ICSI Research Group


ICSI Publication Type

Article in conference proceedings