멀티해상도 헥스평면 융합 가우시안 스플래팅 기반 다중 인체 상호작용 복원

초록

This paper proposes a multi-human interaction reconstruction framework that integrates multi-resolution HexPlane encoding withGaussian Splatting (GS). The proposed architecture constructs HexPlane features that jointly capture global geometry andfine-grained appearance by leveraging vertex normals and pose signals, while the appearance variation module reliably estimatesgeometric and color variations. During training, in addition to the geometry-alignment loss used in GeoAvatar, we introduce anMLP offset regularization loss and a HexPlane grid–based total variation loss, thereby expanding the overall loss structure. Theselosses suppress excessive shape distortion and enhance the spatiotemporal smoothness and consistency of the HexPlane features,achieving a balanced reconstruction of global and local geometry. Experiments on the Hi4D dataset demonstrate that the proposedmethod consistently outperforms Multi-GART [1] and GeoAvatar [2] in key metrics such as PSNR, LPIPS, and P2S, and maintainsstable reconstruction performance even in highly dynamic scenes involving rapid motion and frequent physical interactions. Thisverifies that our approach can faithfully reconstruct complex close-range human interactions with high precision while preservingboundary sharpness and fine details, even when using only eight sparse input views.

키워드