Rand transformer net: An efficient network for semantic segmentation of railway engineering entities based on 3D...

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• 3D point clouds are crucial for autonomous driving, AR, and reconstruction.
• Challenges exist in achieving high accuracy and efficiency in 3D point semantic segmentation due to data complexity.
• Existing methods struggle to balance performance and computational cost for large-scale point cloud applications.

Purpose of the Study:

• To propose a lightweight and efficient network for 3D point semantic segmentation.
• To enhance the capture of local geometric features and address semantic ambiguity in boundary regions.
• To achieve superior performance and scalability for large-scale point cloud scenes.

Main Methods:

• Developed a lightweight Rand Transformer Net (RTN) with an efficient multi-scale feature extraction module.
• Utilized a random downsampling strategy for improved feature extraction.
• Incorporated a Rand Transformer Block for capturing local geometric features.
• Introduced a novel loss function (ABL loss) to constrain labeled boundaries and reduce semantic ambiguity.

Main Results:

• RTN demonstrated superior performance compared to existing methods on the Bridge Dataset.
• The proposed method shows strong scalability and efficiency advantages for large-scale point cloud scenes.
• Experimental results validate the effectiveness of the random downsampling and Rand Transformer Block.

Conclusions:

• The lightweight RTN effectively balances accuracy and computational efficiency for 3D point semantic segmentation.
• The novel ABL loss function successfully addresses semantic ambiguity in boundary regions.
• RTN offers a promising solution for real-world applications involving large-scale 3D point cloud data.