Device and method for selecting training data for 3D object recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing active learning algorithms for 3D object recognition fail to effectively select training data considering both informativeness and diversity, leading to suboptimal performance and high labeling costs.
Innovation Solution
A multi-processor system selects initial training data based on diversity, followed by iterative primary and secondary data selection based on informativeness, and final data selection based on diversity relationships to achieve optimal learning performance with minimal data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If data is selected based solely on informativeness, then the selected data has high informativeness, but the diversity of data is insufficient
Solution Approach 1:
The patent segments the data selection process into three distinct stages performed by separate processors: (1) initial training data selection based on diversity using K-means clustering, (2) training data selection based on informativeness using entropy and inconsistency metrics, and (3) final training data selection based on diversity relationships among selected data. This segmentation allows each processor to specialize in one criterion while the overall system benefits from all three criteria being applied sequentially.
2Adaptability or versatility
If data is selected based solely on diversity, then the diversity of data is high, but the usefulness for improving model performance is reduced
Solution Approach 1:
The first processor performs preliminary data selection based on diversity using K-means clustering to generate an initial training dataset before the second processor applies informativeness-based filtering. This preliminary action ensures that the pool of candidate data already contains diverse representations, which then get refined by subsequent informativeness-based selection rather than being selected directly for informativeness alone.
3Reliability
If a large amount of training data is used, then the learning performance may be improved, but the labeling costs increase
Solution Approach 1:
The patent employs feedback mechanisms where trained neural network models are used to evaluate unlabeled data and provide feedback on their informativeness (through entropy and inconsistency calculations) and diversity relationships. This feedback loop enables the system to identify and select only the most valuable data points for training, achieving high learning performance with minimal training data by iteratively refining selections based on model feedback.
4Productivity
If existing active learning algorithms are used, then data selection is performed, but the validity and usefulness of selected data is low
Solution Approach 1:
The patent combines multiple data selection criteria (diversity, informativeness, and validity) into a composite selection framework where each criterion is applied by a specialized processor in sequence. This composite approach integrates the strengths of different selection methods: diversity-based initial selection, informativeness-based filtering, and validity-based final selection, creating a more robust and effective data selection system than any single method alone.
Data Source
AI summary
Provided is a device for selecting training data for 3D object recognition, which includes a first processor configured to select initial training data based on diversity from an original dataset, a second processor configured to select training data based on informativeness from the initial training data selected by the first processor, and a third processor configured to calculate diversity relationships among the training data selected by the second processor and to select final training data.