An efficient point cloud semantic segmentation network with multiscale super-patch transformer.

Area of Science:

• Computer Vision
• 3D Data Processing
• Deep Learning

Background:

• Semantic segmentation of large-scale 3D point clouds is crucial for environmental perception.
• Challenges include vast data size, efficient deep learning training, and handling object variety/occlusion.

Purpose of the Study:

• To propose an efficient and effective deep learning model for semantic segmentation of large-scale point cloud scenes.
• To address the computational challenges and improve representation of diverse 3D objects.

Main Methods:

• Introduced the Multiscale Super-Patch Transformer Network (MSSPTNet) utilizing scene super-patches as data representation.
• Developed a Multiscale Super-Patch Local Aggregation (MSSPLA) module and a Super-Patch Transformer (SPT) module.
• Employed a dynamic region-growing algorithm for super-patch extraction and self-attention for feature learning.

Main Results:

• MSSPTNet effectively learns both local and global features from point cloud data.
• Demonstrated significant efficiency gains in network training, being tens to hundreds of times faster than existing methods.
• Achieved strong performance on the S3DIS dataset, particularly for indoor scenes with repetitive structures.

Conclusions:

• MSSPTNet offers a highly efficient solution for semantic segmentation of large-scale point cloud scenes.
• The super-patch representation and transformer architecture effectively capture contextual information and inter-patch relationships.
• The proposed method significantly reduces training time, making large-scale 3D scene understanding more accessible.