Face Recognition Using 3D Pose Synthesis for Occlusion Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Digital image recognition systems face challenges in accurately identifying objects and faces in images due to variations in pose and occlusion, leading to degraded recognition performance, especially in consumer photographs where objects are often tilted or partially occluded.
Innovation Solution
The system generates synthesized images to compensate for pose and occlusion by using a three-dimensional model to reorient and reposition objects, allowing for recognition of objects in frontal view, even when they are originally posed at an angle or occluded, and adapts the training set dynamically based on recognition performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If face recognition algorithms are applied to consumer photographs, then recognition capability is enabled, but recognition accuracy degrades due to pose variations and occlusion
Solution Approach 1:
The patent transitions from two-dimensional image analysis to three-dimensional spatial reasoning by inferring 3D pose parameters (rotation and translation) from 2D image data. This dimensional transformation allows the system to understand and compensate for pose variations, effectively resolving the contradiction between maintaining recognition accuracy and tolerating pose variations.
Solution Approach 2:
The system changes the parameter space from raw pixel values to structured pose parameters (rotation angles, translation vectors). By transforming the problem into parameter space, the algorithm can systematically handle pose variations through parameter adjustment rather than being constrained by fixed pose requirements, thus improving both accuracy and adaptability.
2Reliability
If training data is manually labeled, then recognition performance improves, but time and labor requirements increase
Solution Approach 1:
The system performs self-labeling by automatically generating pose labels through its own pose estimation algorithm. Instead of requiring external manual annotation, the algorithm labels training data itself, eliminating the time-consuming manual labeling process while maintaining performance improvement through consistent, algorithm-generated labels.
Solution Approach 2:
The pose estimation algorithm performs preliminary labeling of training data before the main recognition training occurs. This preliminary action prepares the data in advance with accurate pose annotations, enabling subsequent recognition training to proceed efficiently without requiring time-consuming manual intervention.
Data Source
AI summary
Embodiments described herein facilitate or enhance the implementation of image recognition processes which can perform recognition on images to identify objects and/or faces by class or by people.