| Author: | Pang, Kaicheng |
| Title: | Personal stylist : an intelligent evaluation model for personalized styling advice |
| Advisors: | Wong, Wai Keung (SFT) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Recommender systems (Information filtering) Fashion merchandising -- Data processing Human-computer interaction Machine learning Hong Kong Polytechnic University -- Dissertations |
| Department: | School of Fashion and Textiles |
| Pages: | xvi, 183 pages : color illustrations |
| Language: | English |
| Abstract: | An intelligent personal stylist, designed to offer personalized fashion recommendations, is becoming increasingly important in the fashion industry. Meeting the crucial requirements of compatibility and personalization, this technology aims to provide users with aesthetically pleasing and tailored fashion outfits. However, existing systems face limitations when deployed in real-world applications, prompting the need for further advancements in this field. The existing literature exhibits three primary limitations: 1). Most existing fashion compatibility models overlook hierarchical relationships among fashion elements and need more explanatory capabilities; 2). Due to the lack of professional fashion knowledge in current evaluation datasets, the assessment of fashion compatibility models is limited to accuracy rather than aesthetic ability; 3). Existing research on personalized fashion recommendation mainly prioritizes user preferences and social media data, neglecting to incorporate diverse appearances of customers in the recommendation process. This thesis aims to develop an intelligent personal stylist to overcome the above limitations. firstly, a Hierarchical Outfit Network is proposed, featuring a multi-layered structure that captures relations among attributes, items, and outfits. The model learns outfit representation in a bottom-up manner, utilizing the attention mechanism to model feature relationships at each level. A gradient penalty loss is also employed to learn the underlying reasons behind the compatibility prediction. Secondly, a novel evaluation dataset, Aesthetic 100 (A100), is developed to assess the aesthetic capabilities of fashion compatibility models. A100 demonstrates three desirable qualities: 1). Completeness, covering two types of standards in the fashion aesthetic system through two independent aesthetic tests; 2) Reliability, being independent of training data and consistent with major indicators; 3) Explainability, identifying essential fashion aesthetic characteristics to evaluate model performance in more detailed dimensions. Thirdly, a new fashion cognition modeling task is introduced to investigate the relationship between outfits and an individual’s physical attributes. A new dataset is constructed, consisting of 29,352 annotated outfits that indicate physical attribute compatibility. Moreover, a Fashion Convolutional Network is proposed to solve the task, comprising an outfit encoder module that encodes fashion attribute features into an outfit embedding using convolutional layers of various window sizes and a multi-label graph convolutional network module that captures label correlations to learn classifiers for physical attributes. The compatibility score is obtained by applying the classifiers to the outfit embedding. Fourthly, a novel framework, Body-shape-Aware Network, is developed to enhance body-aware recommendations. This network utilizes visual and anthropometric features from a large-scale body shape dataset to represent body shapes. It also incorporates try-on images generated by the proposed Multilayer Virtual Try-on System to represent outfits. The cross-model attention mechanism is leveraged to provide attribute-level explanations. The contributions of this thesis lie in developing and advancing intelligent fashion recommendation systems. The proposed solutions address the challenges of fashion compatibility and cognition modeling, providing practical and effective tools for the fashion industry to enhance customer experiences. The research also reveals its limitations and provides insights for further exploration. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12830