Manifold Dictionary Learning for Person Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for robust person matching across multiple cameras with different viewing angles and lighting conditions are impractical due to the need for extensive manual labor and prior knowledge of matching correspondences, especially in unsupervised learning approaches like dictionary learning, which require positive training images and manual annotation.
Innovation Solution
A method using a modified manifold-based dictionary learning approach that determines correspondences for codes using pairwise similarities between codes, allowing for comparison of objects in images without the need for extensive positive training images or manual annotation, by learning a dictionary from a training dataset collected from multiple cameras.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning methods are used for person matching, then matching accuracy can be improved, but extensive manual labor and prior knowledge of matching correspondences are required
Solution Approach 1:
The system performs self-service by automatically learning the dictionary and code mapping relationships from training images without requiring manual annotation of matching correspondences. The unsupervised learning approach allows the system to autonomously establish the transformation between appearance descriptors and codes, eliminating the need for human annotators to label positive and negative training samples.
Solution Approach 2:
The patent replaces the mechanical process of manual annotation and supervision with an automated computational process. Instead of manually creating labeled training data, the system uses unsupervised dictionary learning algorithms to automatically learn the mapping between appearance descriptors and codes, substituting human labor with computational automation.
2Extent of automation
If dictionary learning is used without positive training images, then manual labor is reduced, but the ability to learn discriminative representations deteriorates
Solution Approach 1:
The system performs preliminary action by pre-defining the dictionary structure and learning objectives before processing training images. The dictionary learning process is initialized with a predefined framework that guides the automatic learning of discriminative representations, ensuring that even without manual annotations, the system can learn meaningful code mappings that preserve matching information.
Solution Approach 2:
The patent changes the parameters of the learning process by using unsupervised dictionary learning instead of supervised methods. This involves modifying the objective function and learning algorithm to work without labeled data, while still achieving discriminative representation through automatic cluster center computation and code assignment based on appearance descriptor similarities.
3Adaptability or versatility
If multiple cameras with different viewing angles and lighting conditions are used, then surveillance coverage is improved, but object appearance variability increases making matching more difficult
Solution Approach 1:
The patent introduces an intermediary transformation process that maps appearance descriptors from different cameras and conditions into a unified code space. The dictionary learning framework acts as an intermediary that learns to represent variations in appearance (due to viewing angles, lighting, etc.) in terms of a fixed set of dictionary atoms, enabling consistent matching across diverse camera conditions.
Solution Approach 2:
The learned dictionary and code mapping serve as a universal representation that works across multiple cameras with different viewing angles and lighting conditions. The system achieves multi-functionality by creating a camera-invariant code space that can represent objects from any camera in the network, allowing the same matching mechanism to work universally across all camera conditions.
Data Source
AI summary
A method of comparing objects in images. A dictionary determined from a plurality of feature vectors formed from a test image and codes formed by applying the dictionary to the feature vectors, is received. The dictionary is based on a modified manifold obtained by determining correspondences for codes using pairwise similarities between codes. Comparison codes are determined for the objects in the images by applying the dictionary to feature vectors of the objects in the images. The objects in the images are compared based on the comparison codes of the objects.