Antialiased 3D Gaussian splatting for human reconstruction: a dual-filter approach with K-NN pruning

This paper proposes an improved 3D Gaussian Splatting (3DGS) framework aimed at enhancing rendering quality and memory efficiency in human-centered 3D scene reconstruction. While 3DGS offers high-quality and real-time rendering, it remains susceptible to aliasing artifacts and excessive memory usage caused by redundant Gaussian elements. To address these challenges, we introduce a dual-stage anti-aliasing mechanism that combines a 3D smoothing filter before projection to limit frequency, and a 2D Mip filter during rendering to simulate the physical image formation process of cameras. Additionally, a k-NN-based redundancy pruning algorithm is proposed to identify and remove overlapping or low-contribution Gaussians, significantly reducing storage requirements without compromising visual fidelity. We conduct extensive experiments on NeRF-Synthetic, Mip-NeRF 360, and a custom dataset built using real-world human captures. Quantitative evaluations using PSNR, SSIM, and LPIPS demonstrate that our method consistently outperforms baseline approaches such as NeRF, Mip-NeRF, and standard 3DGS, especially under lower-resolution training. The pruning strategy further reduces memory consumption by 50–60%, making our framework more suitable for deployment on resource-limited devices. Overall, the proposed enhancements improve both the quality and efficiency of 3DGS, enabling more practical applications in real-time human modeling and immersive environments.