BROWSE
Detail
Conference Paper
3 차원 파운데이션 모델 미세조정을 통한 일반화 가능한 다중객체 3 차원 가우시안 스플래팅
Cited - time in 
Share 
Share
- Issue Date
- 2026-06
- Citation
- 한국방송·미디어공학회 학술 대회 (하계) 2026, pp.1-3
- Publisher
- 한국방송·미디어공학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- Although recent 3D foundation models have shown significant performance improvements, they still suffer from limited performance in human-centered reconstruction due to the domain gap between generic pretraining datasets and human-specific multi-view datasets. In this work, we propose a method for adapting a pretrained 3D foundation model to human-centered multi-view reconstruction via low-rank adaptation. To improve human-specific 3D reconstruction, we further employ a depth loss with ground-truth depth and a Chamfer distance loss between the SMPL vertices and the Gaussian positions during fine-tuning. These complementary supervision signals enhance the reconstruction of human structure in complex multi-person scenes. We validate the proposed method on the Hi4D dataset, which includes diverse multi-person interaction scenarios.
- KSP Keywords
- 3d reconstruction, Chamfer distance, Low Rank, Multi-person interaction, fine-tuning, ground truth, human-centered, multi-view reconstruction, rank adaptation