가우시안 스플래팅 기반 고정밀 3D 수목 에셋 생성 파이프라인
Gaussian Splatting Based Pipeline for High-Fedelity 3D Tree Asset Generation
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Recent advances in 3D data-driven digital twin research have revealed limitations in existing tree reconstruction methods, which rely solely on either scanning or procedural generation. To address this issue, this study proposes a hybrid pipeline that integrates data-driven reconstruction and procedural generation using Gaussian Splatting(GS) data. The proposed method converts multi-view GS outputs into dense point clouds and extracts a stable skeletal structure through color-density-based graph analysis. Fine branches and leaves are procedurally generated using a space colonization algorithm that incorporates botanical principles, achieving a natural and structurally coherent form. Quantitative evaluations using Chamfer distance and Intersection-over-Union metrics demonstrate high geometric similarity and volumetric consistency with the original GS data. The proposed GS-based hybrid framework ensures both visual realism and biological plausibility, enabling efficient and reliable digital twin tree modeling.
Abstract
Recent advances in 3D data-driven digital twin research have revealed limitations in existing tree reconstruction methods, which rely solely on either scanning or procedural generation. To address this issue, this study proposes a hybrid pipeline that integrates data-driven reconstruction and procedural generation using Gaussian Splatting(GS) data. The proposed method converts multi-view GS outputs into dense point clouds and extracts a stable skeletal structure through color-density-based graph analysis. Fine branches and leaves are procedurally generated using a space colonization algorithm that incorporates botanical principles, achieving a natural and structurally coherent form. Quantitative evaluations using Chamfer distance and Intersection-over-Union metrics demonstrate high geometric similarity and volumetric consistency with the original GS data. The proposed GS-based hybrid framework ensures both visual realism and biological plausibility, enabling efficient and reliable digital twin tree modeling.