Dynamic routing towards few-shot point cloud semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Few-shot learning is crucial for 3D point-cloud semantic segmentation to reduce data dependency.
• Existing prototype-based methods often overlook global dependencies and prototype-level context between query and support sets, limiting accuracy.

Purpose of the Study:

• To propose a novel Adaptive-correlated Prototype-oriented dynamic Routing (APR) framework.
• To enhance the accuracy and robustness of few-shot 3D point-cloud semantic segmentation by addressing limitations in current methods.

Main Methods:

• The APR framework employs dynamic routing to explore the context of query-support prototypes.
• Dynamic routing coefficients are used to analyze correlations between support and query prototypes.
• These correlations are integrated into the loss function to improve label prediction accuracy.

Main Results:

• The proposed APR framework demonstrates superior performance on benchmark datasets.
• Experiments on S3DIS and ScanNet datasets validate the effectiveness of the approach.

Conclusions:

• The APR framework effectively addresses the limitations of existing methods by incorporating prototype-level context.
• The method significantly improves the accuracy and robustness of few-shot 3D point-cloud semantic segmentation.