InstantSplat: Sparse-View SfM-Free Gaussian Splatting in Seconds

InstantSplat is a new framework designed to enhance novel view synthesis (NVS) from sparse images without relying on the traditional Structure-from-Motion (SfM) process, which can be time-consuming and unreliable in scenarios with limited matched features. The method integrates dense stereo predictions with point-based representations to create 3D Gaussians from sparse-view data in a matter of seconds. It begins by generating a pixel-wise multi-view stereo dense point cloud and computing initial camera poses. To optimize the representation, InstantSplat employs a grid-based, confidence-aware Farthest Point Sampling technique to strategically position point primitives, minimizing redundancy and ensuring uniform sampling. This approach allows for the initialization of densely populated Gaussians and a streamlined optimization process that adjusts Gaussian and camera parameters efficiently, avoiding the complexities of previous methods. InstantSplat significantly reduces training time from hours to seconds and demonstrates robust performance across various datasets and view counts. The results indicate that this method can effectively handle the challenges posed by sparse-view scenarios, providing a more efficient and reliable solution for 3D scene reconstruction. The framework's performance has been validated against benchmarks, showcasing its advantages over existing techniques.