基于多通道Mamba自适应高斯溅射网络的高精度单视图3D重建

Multi-Channel Mamba Adaptive Gaussian Splatting Network for High-Precision Single-View 3D Reconstruction

摘要: 因缺乏多视角信息，现有单视图三维重建方法普遍难以兼顾建模精度与泛化能力。为此，本文提出一种多通道Mamba自适应高斯溅射网络(MMAGS)来实现高精度重建。该方法在网络结构上引入多通道并行Vision Mamba模块，通过通道划分与并行驱动提升遮挡区域的几何还原能力与跨视角结构信息的稳健提取。为增强纹理细节与图像边缘保真度，进一步提出深度–颜色双梯度感知的自适应三维高斯滤波策略，依据局部几何与纹理变化动态调整高斯协方差。所提出的方法应用于ShapeNet-SRN与CO3D两类数据集。实验结果表明，所提出的MMAGS方法展现出比现有最先进方法更高的重建精度和重建效果，具备良好的跨类别泛化能力与真实场景鲁棒性。

Abstract: Single-view 3D reconstruction faces significant challenges due to the lack of multi-view constraints, often leading to a compromise between modeling accuracy and cross-category generalization. We propose MMAGS, a Multi-channel Mamba Adaptive Gaussian Splatting Network designed for high-precision single-view 3D reconstruction. This framework incorporates a multi-branch Vision-Mamba backbone that segments intermediate features into parallel channels, facilitating robust structural reasoning and improved geometric recovery in occluded and texture-limited areas. To enhance detail preservation further, we introduce a depth-color dual-gradient adaptive 3D Gaussian filter that dynamically adjusts Gaussian covariance according to local geometric and photometric variations, thereby significantly enhancing edge sharpness and texture fidelity. We evaluate MMAGS on the ShapeNet-SRN and CO3D benchmarks, where it demonstrates state-of-the-art reconstruction accuracy and perceptual quality, exhibiting strong cross-category generalization and robustness across both synthetic and real-world datasets.