DANet: Joint Density- and Semantics-Adaptive Convolution for 3D Point-Cloud Semantic Segmentation

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Semantic segmentation of 3D point clouds is challenging due to uneven data sampling from sources like LiDAR.
• Existing methods struggle to adapt to varying point densities and semantic complexities.

Purpose of the Study:

• To introduce DANet, a novel framework for 3D semantic segmentation.
• To address the limitations of uneven sampling in 3D point cloud data.
• To improve segmentation accuracy by adapting to local point density and semantic information.

Main Methods:

• Developed DANet, a framework utilizing joint density- and semantics-adaptive convolution.
• Introduced Density-Adaptive Radius Convolution (DAR-Conv) to predict point-wise neighborhood radii.
• Incorporated Gated Adaptive Cross-Layer Fusion (GACF) for feature alignment and fusion.

Main Results:

• DANet achieved the highest reported mean accuracy (mAcc) on the S3DIS dataset.
• Demonstrated high mean Intersection over Union (mIoU) and overall accuracy (OA) on the NPM3D dataset.
• Validated the effectiveness of density- and semantics-aware receptive-field adaptation.

Conclusions:

• DANet offers a robust solution for 3D semantic segmentation with unevenly sampled data.
• The proposed adaptive convolution approach significantly enhances segmentation performance.
• The framework's ability to adapt receptive fields is crucial for handling diverse point cloud characteristics.