3D Semantic Scene Completion and Occupancy Prediction for Autonomous Driving: A Survey

In recent years, autonomous driving perception technology has rapidly advanced. 3D object detection, as a vital task of perception, has made significant progress. Despite these advancements, there are still various limitations when facing real-world traffic scenarios. However, these issues can be effectively addressed through 3D semantic scene completion and occupancy prediction tasks. Therefore, reviewing existing methods for these two tasks holds significant value. In this comprehensive survey, based on the types of input information, we classify the existing methods into camera-only and LiDAR-only approaches. And we further subdivide them into several subcategories. We introduce the core concepts and specific methods of various approaches within these categories and summarize their advantages and disadvantages. Finally, we provide the performance of these methods on different 3D semantic scene completion and occupancy prediction datasets.