A Novel Point Cloud Encoding Method Based on Local Information for 3D Classification and Segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are robust to point cloud perturbations, crucial for Internet of Things data.
• Current deep learning methods struggle to capture fine-grained local information in point clouds, often increasing network complexity.
• Integrating features from different network levels can yield local information but complicates the model.

Purpose of the Study:

• To propose an effective point cloud encoding method for deep learning networks.
• To enable deep learning models to better utilize local information within point clouds.
• To enhance the recognition of fine-grained features in 3D objects.

Main Methods:

• Developed a point cloud encoding method using an axis-aligned cube to define local regions.
• Constructed feature representations for each point using all points within its local region.
• Input these enhanced feature representations into a deep learning network.

Main Results:

• The proposed method, when integrated with a simple deep learning network, achieved higher accuracy.
• Demonstrated superior performance in 3D object classification on the ModelNet40 benchmark.
• Showcased improved semantic segmentation on the Stanford 3D Indoor Semantics Dataset.

Conclusions:

• The proposed point cloud encoding method effectively captures local information for deep learning.
• Achieved higher accuracy in 3D object classification and semantic segmentation compared to complex methods.
• Offers a simpler yet more effective approach for processing point cloud data in deep learning applications.