3D CNN for Dense Point Cloud Correspondence

The patent segments the feature extraction process into multiple distinct layers: shallow layers (first through third convolutional layers) that extract low-level geometric features with high precision, and deep layers (fourth through sixth convolutional layers) that extract high-level semantic features with high recall. This segmentation allows each layer to specialize in different feature types, resolving the contradiction between precision and recall by combining their outputs in the fusion layer.

The patent merges the outputs of shallow feature extraction layers and deep feature extraction layers in a fusion layer. The shallow layers provide precise low-level geometric features while the deep layers provide robust high-level semantic features. By combining these complementary features through merging, the system achieves both high precision and high recall simultaneously, resolving the contradiction between these two performance metrics.

The patent replaces traditional hand-crafted descriptor methods with a task-specific learned descriptor system based on 3D CNNs. Instead of using manually designed features, the system learns optimal features automatically through training on task-specific data. This substitution of mechanical hand-crafting with automated learning mechanisms achieves superior performance while managing computational complexity through efficient network architecture design.

The patent implements a multi-functional 3D CNN architecture where a single unified network performs multiple functions: shallow layers extract geometric features, deep layers extract semantic features, and the fusion layer integrates both. This universal architecture handles diverse feature extraction tasks within one system, capturing complete information (both geometric and semantic) while managing complexity through shared weights and a structured hierarchical design that reuses computational resources across different feature types.