Institutional Repository of Vilnius University
Investigation of semi-supervised learning methods for face attributes prediction /
Abstract
dc:descriptionFace attributes prediction is used in human identification and verification tasks. However, like most machine learning domains, this one struggles to find enough usable data. This paper analysed semi-supervised learning methods and proposed as well as implemented a method which could infer useful information from massive unlabelled WebFace46M dataset using labelled CelebA dataset. It could help to overcome data deficit issue in face attributes prediction domain. Similarity matrix which was developed in the internship has been used to process additional information between labelled and unlabelled data. Combining this data with proposed attention pooling method this paper managed to extract latent attributes with confidence scores. Using sample from both datasets (8000 labelled, 2000 unlabelled face images) two developed programs were tested. Similarity reading and inversion was performed in 28 seconds. Attribute matrix was created 19 seconds. This research also suggested possible performance optimization for former operation which could have boosted the performance.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Augūnas, Žygimantas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD146236093&prefLang=en_US