Cross-Modality Visual Search via Learned Appearance Mapping

The patent introduces an intermediary mapping system that learns correspondence relationships between RGB and infrared modalities. This mapping acts as a mediator that translates appearance features from one modality to another, enabling cross-modality search without direct complex comparisons. The mapping is learned through training data and can be pruned to remove incorrect mappings, thus reducing false positives while maintaining search coverage.

The patent addresses lighting condition variations by learning mappings that are robust to parameter changes in appearance. The system trains on diverse lighting conditions and learns to map appearances across different lighting scenarios. The mapping can be pruned based on performance metrics to remove mappings that produce false positives under specific lighting conditions, thus maintaining accuracy while preserving robustness.

The patent extracts and retains only the beneficial mappings from the set of all possible mappings. Through training and evaluation, the system identifies which mappings produce accurate results and which produce false positives. Incorrect or redundant mappings are pruned and removed, keeping only the essential mappings needed for effective cross-modality and cross-lighting-condition search. This reduces computational complexity while maintaining search completeness.