Conceitos essenciais
OmniGS introduces a novel omnidirectional Gaussian splatting system that enables fast and high-fidelity reconstruction of radiance fields from calibrated omnidirectional images, outperforming state-of-the-art neural radiance field methods in both reconstruction quality and rendering speed.
Resumo
The paper presents OmniGS, a novel photorealistic reconstruction system that leverages omnidirectional Gaussian splatting for fast and high-quality radiance field reconstruction.
Key highlights:
- Theoretical analysis of the spherical camera model derivatives in 3D Gaussian splatting, enabling direct splatting of 3D Gaussians onto the equirectangular screen space.
- Development of a new GPU-accelerated omnidirectional rasterizer that directly splats 3D Gaussians onto the equirectangular image plane, enabling efficient and differentiable optimization of the radiance field.
- Extensive experiments on egocentric and roaming scenarios demonstrate that OmniGS achieves state-of-the-art reconstruction quality and high rendering speed using omnidirectional images, outperforming NeRF-based methods.
- Evaluation on perspective rendering shows that OmniGS can generate better perspective views by cropping the rendered omnidirectional images, compared to the original 3D Gaussian splatting approach.
The paper highlights the potential of OmniGS for real-time applications in robotics and immersive scene exploration.
Estatísticas
The paper reports the following key metrics:
On the 360Roam dataset, OmniGS achieves a PSNR of 25.505, SSIM of 0.808, and LPIPS of 0.140, with a rendering FPS of 120.
On the EgoNeRF-OmniBlender dataset, OmniGS achieves a PSNR of 33.637, SSIM of 0.919, and LPIPS of 0.054, with a rendering FPS of 115.
On the EgoNeRF-Ricoh360 dataset, OmniGS achieves a PSNR of 26.034, SSIM of 0.825, and LPIPS of 0.131, with a rendering FPS of 93.