This paper introduces a novel point-based representation for 3D occupancy prediction from multi-view images, named *Occupancy as Set of Points* (OSP). Unlike existing camera-based methods that use dense volume-based representations, OSP focuses on *Points of Interest* (PoIs) to represent the scene, allowing for more flexible and adaptable training and inference. The OSP framework consists of an image backbone, a 3D positioning encoder, and a decoder. It excels in handling complex scenes and can be seamlessly integrated with volume-based methods to enhance their effectiveness. Experiments on the Occ3D-nuScenes benchmark demonstrate strong performance and flexibility, achieving a mIoU of 39.4. The method is particularly effective in detecting small targets and can predict areas beyond the standard perception range. The paper also includes ablation studies and a comparison with existing methods, highlighting the benefits of the proposed approach.This paper introduces a novel point-based representation for 3D occupancy prediction from multi-view images, named *Occupancy as Set of Points* (OSP). Unlike existing camera-based methods that use dense volume-based representations, OSP focuses on *Points of Interest* (PoIs) to represent the scene, allowing for more flexible and adaptable training and inference. The OSP framework consists of an image backbone, a 3D positioning encoder, and a decoder. It excels in handling complex scenes and can be seamlessly integrated with volume-based methods to enhance their effectiveness. Experiments on the Occ3D-nuScenes benchmark demonstrate strong performance and flexibility, achieving a mIoU of 39.4. The method is particularly effective in detecting small targets and can predict areas beyond the standard perception range. The paper also includes ablation studies and a comparison with existing methods, highlighting the benefits of the proposed approach.