Semantic Scene Completion With 2D and 3D Feature Fusion

3D semantic scene completion (SSC) aims to get a dense semantic understanding of an environment in 3D. It requires a geometric and semantic knowledge of the surrounding environment and the filling of void areas. In this paper, we propose an improved algorithm by modifying VoxFormer. VoxFormer consists of two steps for 3D semantic scene completion. First, it predicts the occupancy of an environment. Then, it completes the semantic scene completion through a masked autoencoder. It requires separate training for two stages, which can cause a disconnect of information from input to output. We propose an improved VoxFormer algorithm that makes end-to-end training possible by integrating occupancy prediction and scene completion. We use pseudo-LiDAR computed by depth estimation as input of 3D CNN, which generates queries for cross attention with 2D features. This makes the process end-to-end by connecting occupancy prediction and semantic scene completion. Experimental results using SemanticKITTI show improvement in the proposed algorithm.