EIDU-Net: edge-preserved inception DenseGCN U-Net for LiDAR point cloud segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud semantic segmentation is crucial for autonomous driving, scene reconstruction, and human-computer interaction.
• Deep learning methods for point cloud processing often struggle to retain local structural and detailed features due to limitations in exploiting geometric and contextual information.

Purpose of the Study:

• To address the limitations of existing encoder-decoder architectures in point cloud semantic segmentation.
• To propose a novel deep learning model, EIDU-Net, that effectively utilizes both geometric details and high-level features.

Main Methods:

• Introduction of the edge-preserving inception DenseGCN U-Net (EIDU-Net).
• Development of the edge-preserved graph pooling (EGP) layer to retain edge feature information during pooling.
• Implementation of the edge-preserved graph unpooling (EGU) layer for efficient feature graph restoration using retained edge features.

Main Results:

• EIDU-Net demonstrates significant improvements in semantic segmentation tasks.
• The method shows remarkable performance on benchmark datasets like S3DIS and Terracotta Warrior fragments.
• The proposed EGP and EGU layers effectively preserve and restore local structural information.

Conclusions:

• EIDU-Net successfully leverages the complementarity between geometric details and high-level features for enhanced point cloud semantic segmentation.
• The novel pooling and unpooling layers are key to the model's ability to capture detailed features.
• The proposed method offers a robust solution for various 3D data processing applications.