Fast semantic scene completion via two-stage representation

Semantic scene completion, or occupancy prediction, aims to complete and segment environments simultaneously based on incomplete sensor observations. Due to its comprehensive perception capabilities, this technology is becoming a trend in 3D scene understanding. However, semantic scene completion methods require high computational resources, which hinders their real-world application. In this work, we propose an efficient two-stage fast semantic scene completion method called Fast-SSC. In the first stage, a Geometric Completion Network built upon the Hybrid Parallel Dilated Block is designed to extract multi-scale geometric features and complete the scene. In the second stage, a Bird's Eye View network with the Spatial Awareness Enhancement Attention mechanism is employed to achieve fast scene segmentation. Extensive experiments on the SemanticKITTI dataset show that the proposed Fast-SSC achieves state-of-the-art performance. Specifically, the Fast-SSC ranks first among BEV-based methods and operates at 41.8 FPS on NVIDIA GeForce GTX 1080 Ti. These results indicate that our Fast-SSC efficiently utilizes scene information and holds the potential for practical deployment. Code is available at https://github.com/six-wood/Fast-SSC.