Learning Device With Symmetry-Based Feature Point Label Merging

The patent merges multiple unique labels (first labels) that correspond to feature points with congruence or mirror symmetry relations into a single shared label (second label). This consolidation allows training data to be shared across multiple feature point types, effectively increasing the amount of training data available for each label without requiring additional data collection. The inference device uses this label conversion mechanism to improve model accuracy and robustness while avoiding the time-consuming process of collecting separate large datasets for each unique feature point label.

The patent implements a universal inference engine that can handle multiple feature point types through label conversion. Instead of preparing separate extractors for each unique label, the system uses a single inference engine that converts first labels to second labels based on congruence or mirror symmetry relations. This allows one extractor to perform the function of multiple extractors would otherwise be needed, significantly reducing preparation labor while maintaining extraction accuracy through the learned relationships between different feature point types.