4D-GAIA:Depth-Guided 4D Gaussian Representation for Dynamic Visual Data Analytics

Abstract

Dynamic visual data, such as monocular videos of moving humans and changing environments, contain rich spatio-temporal information but are often large, redundant, and difficult to represent in a compact and structured form. 4D Gaussian Splatting has recently emerged as an explicit representation for dynamic scene reconstruction, enabling efficient rendering of time-varying visual content. However, monocular dynamic reconstruction remains challenging due to depth ambiguity, unstable geometry, and weak supervision in regions with fast motion or appearance variation. To address this issue, this paper presents a depth-guided 4D Gaussian Splatting framework for compact representation of dynamic visual data. The proposed framework incorporates monocular depth priors into the 4DGS optimization process to provide additional geometric supervision beyond the RGB reconstruction loss. We study how depth regularization affects the reconstruction stability, visual quality, and spatio-temporal representation of dynamic scenes. Experimental results under a monocular dynamic video setting show that depth guidance improves geometric consistency and helps stabilize scene representation. The proposed framework provides a practical step toward depth-aware dynamic visual data representation for visual analytics, digital twins, and intelligent scene understanding.