Enhancing 3D Asset Retrieval with Semantic Search

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2024-05-20

Department

Major/Subject

Computer Science

Mcode

SCI3042

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

45

Series

Abstract

Semantic search goes beyond exact keyword matching and instead focuses on the comprehension of search queries' purpose and the surrounding context to deliver relevant search results. In this work, we focus on integrating hybrid search, which uses semantic search and traditional lexical search together, in a 3D asset retrieval pipeline of a game engine. We use the embedding of the 3D asset's rendering, title, and description to fill our vector search index for the semantic search part. This allows our search pipeline to use the visual embedding of the asset for search in case of missing or inaccurate metadata. We rely on the recent advances in Deep Learning that have led to the training of capable multimodal models that can embed various data modalities in the same embedding space. In this work, we mainly use vision-language models that embed images and text in the same embedding space. We test our proposed hybrid search method on a dataset of 5000+ assets provided and labeled by HypeHype. We demonstrate the benefits of our method compared to a pure lexical search method through an A/B test and report the improvements in click-through rate and average relevant index of the search results.

Description

Supervisor

Hämäläinen, Perttu

Thesis advisor

Relas, Asko

Keywords

information retrieval, semantic search, hybrid search, multimodal deep learning, embeddings, vector search

Other note

Citation