Vehicle Image Selection for Deep Learning Teacher Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection techniques using deep neural networks face challenges in selecting suitable images as teacher data for learning, especially when computational resources are limited and real-time performance is required, as they require significant memory or continuous data transmission, making it impractical to record all in-vehicle camera images.
Innovation Solution
An image selection device and method that tracks objects across a series of images using a classifier, selecting images based on tracking duration and object size thresholds, as well as confidence scores, to identify images where object detection may have failed, thereby providing suitable teacher data for improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all images generated by the in-vehicle camera are recorded to acquire image data as teacher data, then detection accuracy for learning is improved, but memory capacity requirements and device complexity increase significantly
Solution Approach 1:
The patent extracts only the necessary images for training by identifying and selecting specific images based on tracking duration and object size criteria, rather than recording all images. This extraction approach reduces the data volume to be stored while maintaining training effectiveness.
Solution Approach 2:
Instead of processing all captured images (excessive action), the system applies partial action by selectively processing only those images that meet the predefined criteria for being useful training data, thereby reducing computational and storage resources required.
2Measurement precision
If all images are recorded to acquire sufficient teacher data, then detection accuracy is improved, but computational resources and transmission requirements increase
Solution Approach 1:
The system extracts and transmits only the selected training images that meet the tracking and size criteria, rather than transmitting all captured images. This significantly reduces the data transmission volume and associated computational resources.
3Speed
If real-time performance is demanded, then system responsiveness is improved, but the ability to process and select training data is reduced
Solution Approach 1:
The system performs preliminary actions during real-time operation by continuously tracking objects and identifying candidate images that meet the criteria. This preliminary identification allows the system to prepare training data selection in advance without compromising real-time responsiveness.
Data Source
AI summary
An image selection device includes a processor configured to: input, for each of a series of images acquired from a camera mounted on a vehicle, the image to a classifier to detect a region including an object represented on the image; track the detected object over the series of images; and select, when a period in which the detected object can be tracked is equal to or more than a predetermined period, and a size of a region including the detected object in any one image during the period in which the object can be tracked is equal to or more than a predetermined size threshold value, among the series of images, an image immediately before the period in which the object can be tracked, or an image in which the tracked object is not represented during the period in which the object can be tracked.