Domain Adaptation in Computer Vision Applications by Gabriela Csurka

Author:Gabriela Csurka
Language: eng
Format: epub, pdf
Publisher: Springer International Publishing, Cham

Other works have contemporaneously explored the idea of directly optimizing a representation for domain invariance [182, 309]. However, they either use weaker measures of domain invariance or make use of optimization methods that are less robust than our proposed method, and they do not attempt to solve the task transfer problem in the semi-supervised setting.

9.2 Joint CNN Architecture for Domain and Task Transfer

We first give an overview of our convolutional network (CNN) architecture, depicted in Fig. 9.2, that learns a representation which both aligns visual domains and transfers the semantic structure from a well-labeled source domain to the sparsely labeled target domain. We assume access to a limited amount of labeled target data, potentially from only a subset of the categories of interest. With limited labels on a subset of the categories, the traditional domain transfer approach of fine-tuning on the available target data [192, 236, 418] is not effective. Instead, since the source labeled data shares the label space of our target domain, we use the source data to guide training of the corresponding classifiers.

Our method takes as input the labeled source data (blue box in Fig. 9.2) and the target data (green box in Fig. 9.2), where the labels are only provided for a subset of the target examples. Our goal is to produce a category classifier that operates on an image feature representation parameterized by representation parameters and can correctly classify target examples at test time.

For a setting with K categories, let our desired classification objective be defined as the standard softmax loss

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.