UniSHARP: Single-Image Novel View Synthesis Across Perspective, Fisheye, and Panoramic Cameras

Researchers from Insta360 Research, Wuhan University, and collaborating institutions have introduced UniSHARP, a monocular novel view synthesis method capable of handling diverse camera types — perspective, wide field-of-view, fisheye, and panoramic — from a single input image. The system predicts a 3D Gaussian point cloud by estimating ray-distance geometry and multi-scale features, then renders high-quality novel views using a unified Gaussian representation across all supported projection types.

Alongside the model, the team releases OmniRooms, a new panoramic simulation dataset built via AirSim comprising 16 large indoor scenes, 300,000 RGB images with depth information, and an accompanying wide-FoV fisheye variant. The dataset and a field-of-view stratified benchmark are designed to provide a standardized testbed for evaluating universal-camera novel view synthesis, an area where no unified evaluation protocol previously existed.