EGNet: 3D Semantic Segmentation Through Point-Voxel-Mesh Data for Euclidean-Geodesic Feature Fusion

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Service robot advancement necessitates higher precision in indoor semantic segmentation.
• Traditional methods using Euclidean features from point cloud/voxel data lack geodesic information, reducing boundary accuracy and increasing computational cost.

Purpose of the Study:

• To propose a novel network, the Euclidean-geodesic network (EGNet), for enhanced indoor semantic segmentation.
• To improve boundary accuracy and computational efficiency by integrating Euclidean and geodesic features.

Main Methods:

• EGNet utilizes point cloud-voxel-mesh data for detail, contour, and geodesic feature characterization.
• Feature fusion occurs via parallel Euclidean and geodesic branches, with geodesic features extracted from mesh data.
• Inter-domain fusion and aggregation modules enhance geodesic feature extraction, combined with Euclidean contextual features.

Main Results:

• Visual comparisons on Scannet and Matterport datasets demonstrate EGNet's effectiveness against other models.
• The integration of Euclidean and geodesic features significantly improves semantic segmentation accuracy.
• EGNet shows superior performance in characterizing object boundaries and details.

Conclusions:

• The proposed EGNet effectively enhances indoor semantic segmentation accuracy by fusing Euclidean and geodesic features.
• This approach offers a promising direction for future research in combining diverse feature types for segmentation tasks.
• EGNet provides a foundation for more robust and precise scene understanding in service robotics.