CNN을 활용한 sLLM 기반의 한국어 텍스트-모션 검색 시스템
sLLM-based Korean Text-to-Motion Retrieval System Using CNN
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
In this paper, we propose a novel approach to improve the performance of Korean text- based 3D human motion retrieval. Previous studies have failed to adequately cover diverse motion representations and have limitations in distinguishing subtle differences between motions. To address these issues, we introduce sLLM to embed comprehensive motion descriptions and utilize CNN to distinguish subtle motion variations. In addition, we created a Korean HumanML3D dataset to adapt previous researches and conduct experiments on Korean. We also developed a new testing protocol and performed additional experiments. The results show that the proposed method outperforms previous researches, with the R@1 metric improving from 4.49 to 6.39 and the MedR metric improving from 35.00 to 24.75. Furthermore, we compared qualitative results and demonstrated that our method performs well.
Abstract
In this paper, we propose a novel approach to improve the performance of Korean text- based 3D human motion retrieval. Previous studies have failed to adequately cover diverse motion representations and have limitations in distinguishing subtle differences between motions. To address these issues, we introduce sLLM to embed comprehensive motion descriptions and utilize CNN to distinguish subtle motion variations. In addition, we created a Korean HumanML3D dataset to adapt previous researches and conduct experiments on Korean. We also developed a new testing protocol and performed additional experiments. The results show that the proposed method outperforms previous researches, with the R@1 metric improving from 4.49 to 6.39 and the MedR metric improving from 35.00 to 24.75. Furthermore, we compared qualitative results and demonstrated that our method performs well.