Social Relation Recognition Network for Group Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision techniques fail to accurately identify the main group of people in an image, as they rely on insufficient spatial cues, leading to inaccuracies in determining group membership and social relationships.
Innovation Solution
A system and method using a Social Relation Recognition Network (SRRN) that combines First Glance and Graph Reasoning models to identify social relationships between individuals in an image, generating a prediction of group membership by processing people data to determine main and secondary persons and their relations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If spatial cues (proximity, relative poses, actions, interactions, trajectories) are used for group identification, then the device complexity is reduced, but the measurement precision of group membership determination deteriorates
Solution Approach 1:
The system segments the group identification task into multiple components: detecting individual people, determining their poses and actions, analyzing spatial relationships, and inferring social connections. This segmentation allows each component to be optimized independently while improving overall accuracy without proportionally increasing complexity.
Solution Approach 2:
The system transitions from analyzing only spatial dimensions (proximity, position) to incorporating temporal dimensions (trajectories, actions over time) and semantic dimensions (pose interpretation, interaction types). This multi-dimensional approach enriches the feature space for group identification, improving precision while distributing complexity across multiple analysis layers.
2Ease of operation
If simple spatial cues are used for group identification, then the ease of operation is improved, but the reliability of group membership determination deteriorates
Solution Approach 1:
The system introduces intermediate features as mediators between simple spatial cues and final group identification. These intermediates include detected poses, actions, interaction types, and temporal patterns, which translate raw spatial data into meaningful social relationship indicators, improving reliability while maintaining operational simplicity.
Solution Approach 2:
The system replaces simple mechanical spatial thresholding with machine learning-based interpretation of poses, actions, and interactions. This substitution enables the system to understand contextual meaning in spatial relationships, significantly improving reliability while the high-level interface remains simple to operate.
3Measurement precision
If machine learning models (First Glance and Graph Reasoning) are combined for social relation recognition, then the measurement precision of group identification is improved, but the device complexity increases
Solution Approach 1:
The system merges two specialized machine learning models (First Glance for initial social relationship detection and Graph Reasoning for relationship inference and group structure analysis) into a unified pipeline. This combination leverages the strengths of each model to achieve higher precision in group identification while managing complexity through integrated architecture design.
Solution Approach 2:
The First Glance model performs preliminary action by quickly identifying potential social relationships and candidate group members before the more computationally intensive Graph Reasoning model refines the analysis. This preliminary filtering reduces the input complexity for the second model while maintaining high overall precision.
Data Source
AI summary
Systems, methods, and computer-readable media for identifying a main group of people in an image via social relation recognition. The main group of people is identified within an image by identifying social relationships between people visible in the image. The identification of social relationships is performed by a Social Relation Recognition Network (SRRN) trained using deep learning. The SRRN combines two techniques for group identification, First Glance and Graph Reasoning, and fuses their outputs to generate a prediction of group membership. A group refinement module improves and filters the group membership after identification of an initial main group.