Thesis for the degree of Master of Science
A Query-Free Approach to 3D Instance Recognition
Using Vision-Language Models
Jeonghan Lee
Department of Information Convergence Engineering
The Graduate School
Pusan National University
February 2027
Overview
Existing open-vocabulary 3D recognition methods can recognize object categories beyond those defined during training, but often still require users to specify target classes in advance at inference time. This thesis investigates open-ended 3D recognition for inferring the semantic identities of 3D instances without predefined class queries. The proposed approach leverages visual and semantic information from vision-language models together with 3D instance representations to perform instance-level recognition. Experiments on indoor 3D datasets evaluate the recognition performance and applicability of the proposed approach under diverse conditions.
Research
Content will be updated.
Publications
Content will be updated.