Institutional Repository of Vilnius University
Text-based fashion search system for the lithuanian language using computer vision /
Abstract
dc:descriptionIn the research paper, the problem of retrieving the most relevant fashion item photo from a Lithuanian textual query is analyzed for a fashion retrieval system. The main goal of this study is to compare the multilingual text encoder performance with the language specific text encoder performance in a Lithuanian retrieval system based on CLIP architecture. An English fashion data set FashionGen was translated into Lithuanian using OpenAi GPT4o mini and prompting techniques (role-based prompting and few-shot prompting with in-domain examples). Two models that differ in text encoders (XLM-Roberta-Large-ViT-L-14 and LitLatBERT-ViT-L-14) were fine-tuned using the symmetric cross entropy loss function on the FashionGen dataset. The fine-tuned models were assessed on retrieval metrics (Recall@K, Precision@K, mean and median rank) and embedding alignment measures (diagonal dominance, accuracy of alignment, contrast ratio). The results show that the fine-tuned XLM-Roberta-Large-ViT-L-14 had more than 60% improvements in the Recall@K and Precision@K scores and small percentage improvements in alignment measure scores that ranged from 0.27% to 2.34%, compared to the base CLIP model, justifying the finetuning process. Meanwhile, LitLatBert embedding space has collapsed into a dense cluster, resulting in highly similar embeddings for both matching and nonmatching text-image pairs which led to a significant drop in retrieval performance.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Čiurlionytė, Gabrielė,
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:ELABAETD239030143&prefLang=en_US