This paper addresses the problem of visual 3D semantic scene completion (SSC), which infers the complete 3D scene layout and semantics from a single 2D image. To overcome the limitations of existing monocular SSC methods that cannot sufficiently cover real-world traffic situations where a significant portion of the scene is occluded or out of the camera view, this paper proposes Creating the Future SSC (CF-SSC), a novel temporal SSC framework that extends the effective perceptual range of the model by leveraging pseudo-future frame prediction. CF-SSC establishes accurate 3D correspondences by combining pose and depth, and geometrically consistently fuses past, present, and predicted future frames in 3D space. Unlike existing methods that rely on simple feature stacking, our 3D perception architecture explicitly models the spatiotemporal relationships to achieve more robust scene completion. We demonstrate state-of-the-art performance through comprehensive experiments on SemanticKITTI and SSCBench-KITTI-360 benchmarks, validating the effectiveness of our method in occluded part inference and improving 3D scene completion accuracy.
