Glossy-Gaussian: Adaptive Anisotropic Gaussians for View-Dependent Appearances

3D Gaussian Splatting (3DGS) has emerged as a groundbreaking technique in the realm of photorealistic and real-time novel view rendering, garnering significant attention across diverse disciplines, including photogrammetry, remote sensing, and computer vision. Despite its remarkable capabilities, 3DGS encounters substantial challenges when applied to scenarios involving glossy or reflective materials, which exhibit pronounced variations in appearance from different viewing angles. To address this limitation, we introduce Glossy-Gaussian, a plug-and-play integration based on 3DGS for glossy scene reconstruction with real-time synthesis. Specifically, we enhance the view-adaptive spherical harmonic (SH)-based Gaussian kernel to introduce anisotropy to the transparency field, enabling the Gaussian field to accurately represent smooth and specular reflective regions without compromising rendering speed. Furthermore, we extend the capabilities of this Gaussian kernel to incorporate rotation and scaling, facilitating the directional decoding of both properties at the same Gaussian position. Our experimental results demonstrate that our approach significantly improves rendering quality in glossy scenes, as evidenced by increases in PSNR, SSIM, and LPIPS values, and achieves state-of-the-art performance compared to existing methods.