Multi-Camera Training Data Selection for Stable Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for obtaining training data for machine learning rely heavily on the precision of initial sensors, making the data unstable and costly, and fail to effectively utilize diverse image capture conditions.
Innovation Solution
An information processing method that selects training data by analyzing the positions and capturing directions of multiple image capturing devices, assigning higher values to object detection results based on device performance, and choosing images with diverse detection outcomes to stabilize the data acquisition process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the precision of object detection by sensors is improved, then the quality of training data is improved, but the cost and complexity of the system increases
Solution Approach 1:
The patent combines multiple image capturing devices with different capture performances into a unified training data selection system. By merging diverse image sources and applying consistent selection criteria across them, the system achieves reliable training data without requiring any single sensor to have high precision, thus resolving the contradiction between detection precision and system complexity
Solution Approach 2:
The patent changes the selection parameter from relying on high detection precision to relying on diversity of capture performance. By evaluating images based on whether they meet minimum thresholds and selecting diverse samples rather than high-precision samples, the system obtains stable training data while reducing the need for expensive high-precision sensors
2Measurement precision
If high-precision sensors are used to ensure training data quality, then the detection accuracy is improved, but the cost increases
Solution Approach 1:
The patent uses multiple standard image capturing devices instead of expensive high-precision sensors. By employing readily available, lower-cost devices and compensating through quantity and diversity rather than individual quality, the system reduces overall system cost while maintaining training data quality through the selection process
Solution Approach 2:
The patent merges multiple standard-grade image capturing devices to achieve the quality that would otherwise require a single high-precision sensor. By combining diverse inputs from multiple standard devices and applying selection criteria, the system achieves cost-effective training data acquisition
3Adaptability or versatility
If multiple image capturing devices with different performances are used, then the diversity of training data is improved, but the difficulty of selecting appropriate data increases
Solution Approach 1:
The patent applies different evaluation criteria to different image capturing devices based on their capture performances. By tailoring the selection approach to each device's characteristics while maintaining consistent quality thresholds, the system manages the complexity of selecting from diverse sources without requiring a completely uniform evaluation method
Solution Approach 2:
The patent changes the selection approach from evaluating individual image quality to evaluating diversity across multiple devices. By selecting images that represent different capture performances rather than focusing on the best individual images, the system simplifies the selection process while achieving diverse training data
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing method includes the following performed using a computer: acquiring images created through capturing at the same time, and positions and capturing directions of image capturing devices that created respectively each of the images; acquiring object detection results obtained respectively through object detection processes performed using each of the images; and performing a training data selection process using the images, the positions and the capturing directions, and the object detection results acquired. The training data selection process includes: calculating a common region that is seen in the images in common, based on the positions and the capturing directions; determining a degree of agreement among the object detection results in the common region; and selecting an image to be used as training data from among the images, according to the degree of agreement.