Key techniques for three-dimensional completion：a review

作者：Xiao, Haihong; Wu, Qiuxia; Li, Yuqiong; Kang, Wenxiong^*

来源：Optics and Precision Engineering, 2023, 31(5): 667-696.

DOI：10.37188/OPE.20233105.0667

• 3-D shape
• 3-D vision
• 3D scenes
• Completion methods
• Depthmap
• Partial observation
• Scene completion
• Semantic scene completion
• Shape completions
• Three dimensional (3D) shapes

The inference of complete three-dimensional（3D）shape and semantic scene information from partial observations is crucial for various applications，such as autonomous driving，robotic vision，and metaverse ecosystem construction. Research on 3D completion has primarily focused on 3D-shape，3D-scene，and 3D-semantic scene completion. In this paper，we systematically summarize and analyze recent relevant studies concerning these 3D completion tasks. First，for 3D-shape completion，the research progress is reviewed from two aspects：traditional shape completion and deep learning-based shape completion. Second，for 3D-scene completion，the research progress is reviewed from two aspects：the scene completion method based on model fitting and the scene completion method based on a generative approach. For 3D-semantic scene completion，the coupling characteristics between the two tasks of scene completion and semantic segmentation are analyzed，and the research progress is reviewed from three aspects：the depth map-based semantic scene completion method，the depth map-based semantic scene completion method with color images，and the point cloud-based semantic scene completion method，according to the different forms of input data. Finally，we analyze the current problems and future development trends of 3D completion tasks，aiming to provide a reference for related studies in this emerging field in 3D vision.
? 2023 Chinese Academy of Sciences.

• 单位
  华南理工大学