Geo-SelfSSC: Integrating Dense Geometric Priors for Enhanced Self-Supervised Semantic Scene Completion

Accurate 3D scene understanding is vital for applications like autonomous driving and robotics. However, existing voxel-based methods struggle with the reliance on large-scale labeled data, inherent voxel-pixel misalignments, and high computational costs. Recent self-supervised Semantic Scene Completion (SSC) methods using Neural Radiance Fields (NeRF) reduce 3D annotation needs but assume Lambertian surfaces and rely on photometric consistency, which fails in real-world scenes with non-Lambertian effects and sparse camera coverage. In this paper, we introduce Geo-SelfSSC, a self-supervised framework that leverages temporal information from consecutive frames and slight variations in camera poses for supervision, while integrating dense geometric cues to enhance the reconstruction quality and efficiency of neural implicit models. Specifically, Geo-SelfSSC leverages depth priors along with complementary semi-local and local geometric supervision to facilitate efficient sampling, while ensuring effective complementarity between photometric and geometric cues. Our method yields significant improvements in reconstructing reflective and under-observed regions where conventional photometric-based strategies struggle. Comprehensive experiments on challenging tasks demonstrate that Geo-SelfSSC not only achieves strong results in semantic scene completion but also establishes competitive performance in geometry-only 3D occupancy prediction and monocular depth estimation. Our code and models are available at: https://github.com/Xidian AIGroup190726/GeoSelfSSC.