KNOWLEDGE EXTERNALIZATION FROM IMAGE-TO-IMAGE TRANSLATION
Abstract
Image-to-image translation based on generative adversarial networks has seen remarkable progress, even with little to no label information associated with training data. This means that these generative models can properly learn rich knowledge required to generate realistic images. For instance, the fact that a generative model can translate a person’s facial image by properly changing the hair color from black to blond indicates that the model can identify the hair region of a given facial image, even without specific segmentation labels indicating the hair region. When a generative model changes a facial expression from frowning to smiling, it should be able to identify the facial parts such as one’s eyes and mouth, even with no specific labels for them. Motivated by such superior capabilities of generative models, our research aims at externalizing the knowledge learned by the generative model, thus making it possible to perform various downstream tasks, such as landmark detection and pose estimation, which typically require a large amount of labels, with a minimal amount of labelling efforts.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 11, 2021
- Source ID
- FA23862014044
Entities
People
- Jaegul Choo
Organizations
- Air Force Office of Scientific Research
- KAIST
- United States Air Force