Few-shot Pose-free View Synthesis with 3D Gaussian Splatting

Recent advancements in Novel View Synthesis using neural radiance fields and 3D Gaussian splatting have demon-strated promising results. However, these methods conventionally rely on accurate initial poses and point clouds from the model of structure from motion, which poses challenges in sparse input views. To addressing these drawbacks, we introduces an novel method for view synthesis based on 3D Gaussian splatting. Our approach operates on sparse input views without the prerequisite of pose priors, achieving reconstruction and real-time rendering of novel perspectives. For the deficiency of camera poses and point cloud, we employ an end- to-end deep learning model to recover the 3D structure from a few of 2D images to initialize the 3D Gaussian point cloud. In addressing the issue with few-shot, we utilize the differentiable rasterization module to render depth map and apply regulation on it. Additionally, we incorporate supervision between input views by interpolating to generate novel perspectives during training. Experimental results demonstrate that our proposed method significantly enhances the capabilities of 3D Gaussian splatting for novel view synthesis in sparse views and without the necessity of camera pose priors.