Regular Paper Accepted at the Outstanding International Conference IEEE/CVF CVPR (The IEEE/CVF Computer Vision and Pattern Recognition conference) 2025

Figure 1. (Left) Lee Su-hyun (first author), graduate; (right) Kim Seo-yeon (co-author), master's researcher

The paper 'GeoAvatar: Geometrically-Consistent Multi-Person Avatar Reconstruction from Sparse Multi-View Videos', submitted by the research team of Professor Lee Ju-ho of the Department of Computer Science & Engineering at Sogang University, was accepted as a regular paper at IEEE/CVF CVPR (The IEEE/CVF Computer Vision and Pattern Recognition conference) 2025, held in Nashville, United States, from 11 to 15 June 2025. IEEE/CVF CVPR is the world's leading conference in computer vision and in computer science.

Figure 2. Overview of GeoAvatar

GeoAvatar proposes a new method for geometrically consistent multi-person avatar reconstruction from sparse multi-view videos. Existing approaches that reconstruct avatars independently fail to reproduce the geometric relationships between people properly, causing interpenetration between avatars. This research uses a 2D Gaussian splatting (2DGS) based avatar representation to generate geometrically accurate surfaces for multiple people, and exploits monocular priors to achieve surface alignment and geometric refinement in textureless regions. In particular, it resolves surface interpenetration efficiently through surface ordering, without explicitly modeling physical interaction. Experiments on the Hi4D dataset show that the proposed method achieves state-of-the-art performance in both rendering quality and geometric accuracy, enabling geometrically consistent avatar reconstruction of multiple people in close interaction from sparse multiple views within a short training time.

The research, carried out jointly with the Electronics and Telecommunications Research Institute (ETRI), was conducted under the supervision of Professor Lee Ju-ho by Lee Su-hyun (first author), a graduate, and Kim Seo-yeon (co-author), a master's researcher, both of Professor Lee's research team. It will be presented on 15 June at IEEE/CVF CVPR 2025, held in Nashville, United States, from 11 to 15 June 2025.

Professor Lee Ju-ho, the paper's corresponding author, said: "As a technique for rapidly reconstructing geometrically consistent multi-person avatars from sparse multi-view videos, it enables 3D-based image processing after filming — changing the camera viewpoint, editing avatars and so on — so it is expected to be used in a variety of applications, including the production of immersive content in virtual reality environments and the creation of digital humans that interact in real time."

Homepage of Professor Lee Ju-ho's research team: https://sites.google.com/view/vclabsg