Object Identification Model Re-learning with Multi-view Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Identification systems face low accuracy in recognizing objects when the training data is biased towards specific directions or views, leading to poor performance when encountering unfamiliar orientations, such as vehicles traveling in opposite directions.
Innovation Solution
The system employs a model re-learning method that utilizes both internal and external generation models to improve identification accuracy by re-training with data labeled based on identification results from different models, integrating reliabilities weighted by similarity in capturing environment attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If training data is collected from a single camera location with fixed orientation, then data collection is simple and fast, but identification accuracy deteriorates for objects in different directions or views
Solution Approach 1:
The patent combines training data from multiple cameras located at different positions and orientations. By merging datasets from multiple sources, the system creates a comprehensive training set that covers various viewing angles and directions, thereby improving identification accuracy for objects in different orientations while maintaining efficient data collection through automated multi-camera deployment
Solution Approach 2:
The patent introduces spatial dimensionality by deploying cameras at multiple locations and orientations rather than relying on a single viewpoint. This multi-dimensional data collection approach ensures that training data encompasses objects from various angles and directions, resolving the contradiction between simple data collection and accurate identification across different views
2Ease of operation
If a single model is used for identification, then the system is simple and fast to operate, but identification accuracy deteriorates when encountering unfamiliar orientations or views
Solution Approach 1:
The patent segments the identification task by employing multiple specialized models, each trained on data from specific camera orientations and locations. Instead of using one general model, the system divides the identification function across multiple models that each excel at recognizing objects from their trained perspectives, thereby maintaining operational simplicity while improving accuracy for diverse views
Solution Approach 2:
The patent creates a multi-functional identification system where multiple models work together to handle various identification scenarios. Each model serves a specific function based on its training data from particular camera viewpoints, but collectively they provide universal identification capability across all orientations and views, resolving the contradiction between simplicity and adaptability
3Productivity
If training data is biased towards specific directions, then data collection is efficient and focused, but identification accuracy deteriorates for objects in opposite or different directions
Solution Approach 1:
The patent applies local quality by having different cameras collect data with different local characteristics (specific orientations and viewpoints). Each camera focuses on capturing objects from its particular direction, creating locally optimized training data that when combined, provides comprehensive coverage of all possible object orientations while maintaining efficient targeted data collection at each location
Data Source
AI summary
Learning means 701 learns a model for identifying an object indicated by data by using training data. First identification means 702 identifies the object indicated by the data by using the model learned by the learning means 701. Second identification means 703 identifies the object indicated by the data as an identification target used by the first identification means 702 by using a model different from the model learned by the learning means 701. The learning means 701 re-learns the model by using the training data including the label for the data determined based on the identification result derived by the second identification means 703 and the data.