Projectpage of GPS-Gaussian
Summary
We current a brand new method, termed GPS-Gaussian, for synthesizing novel views of a personality in a real-time method. The proposed methodology permits 2K-resolution rendering beneath a sparse-view digicam setting. In contrast to the unique Gaussian Splatting or neural implicit rendering strategies that necessitate per-subject optimizations, we introduce Gaussian parameter maps outlined on the supply views and regress straight Gaussian Splatting properties for immediate novel view synthesis with none fine-tuning or optimization. To this finish, we practice our Gaussian parameter regression module on a considerable amount of human scan information, collectively with a depth estimation module to raise 2D parameter maps to 3D house. The proposed framework is totally differentiable and experiments on a number of datasets display that our methodology outperforms state-of-the-art strategies whereas attaining an exceeding rendering velocity.
