Curation Bias in Domain Adaptation

Distribution Rights granted to UMBC by the author.
This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see http://aok.lib.umbc.edu/specoll/repro.php or contact Special Collections at speccoll(at)umbc.edu

Abstract

Domain adaptation is an important problem with many practical applications. The goal is to adapt a model trained on one domain (source) to another domain (target) with scarce or no annotation. We observe that the unlabeled target datasets of popular domain adaptation benchmarks do not contain any categories apart from testing categories. We believe this introduces a bias that does not exist in many practical applications. We note that this bias can be reduced easily by appending the datasets with images from non-testing categories. On these modified benchmarks, state-of-the-art domain adaptation methods show a large drop in performance. Thus, raising concerns about their practical applicability. Further, we show that a simple, two-stage method involving self-supervised task of rotation prediction and knowledge distillation is a competitive baseline.

Curation Bias in Domain Adaptation

Files

Links to Files

Permanent Link

Collections

Author/Creator

Author/Creator ORCID

Date

Type of Work

Department

Program

Citation of Original Publication

Rights

Subjects

Abstract