USPLAT4D
收藏资源简介:
USPLAT4D是一个基于单目视频的动态高斯散点模型,用于4D场景重建。该模型通过引入不确定性感知机制,能够有效地处理遮挡和极端视角下的动态场景重建问题。USPLAT4D通过评估每个高斯散点的可靠性,构建了一个时空图,并通过这个图来传播可靠的动态线索,从而提高动态场景重建的稳定性和质量。实验表明,USPLAT4D在多种真实和合成数据集上均取得了显著的性能提升,尤其是在极端视角下,能够生成更加准确和一致的4D重建结果。
USPLAT4D is a monocular video-based dynamic Gaussian splatting model designed for 4D scene reconstruction. This model introduces an uncertainty-aware mechanism to effectively address dynamic scene reconstruction challenges under occlusions and extreme viewpoints. USPLAT4D constructs a spatiotemporal graph by evaluating the reliability of each Gaussian splat, and propagates reliable dynamic cues via this graph to enhance the stability and quality of dynamic scene reconstruction. Experimental results demonstrate that USPLAT4D achieves significant performance improvements across various real and synthetic datasets, and particularly generates more accurate and consistent 4D reconstruction results under extreme viewpoints.

- 1Uncertainty Matters in Dynamic Gaussian Splatting for Monocular 4D Reconstruction德克萨斯A&M大学、梅赛德斯-奔驰北美公司 · 2025年



