Automated Image Labeling for Autonomous Vehicle Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for selecting training data for image learning in autonomous vehicles are labor-intensive and time-consuming, especially when identifying images with occlusive objects, which are rare and require manual selection.
Innovation Solution
An apparatus and method that utilize an interest network and an auxiliary network to automatically recognize objects of interest, detect occlusive objects, and determine the reliability scores for image segmentation and object detection, thereby deciding whether to label the image for training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual selection of images containing occlusive objects is performed, then the training data quality for improving object recognition performance is enhanced, but the labor and time required for data selection increases significantly
Solution Approach 1:
The patent replaces the mechanical manual selection process with an automated computational system. The apparatus uses an interest network to detect objects of interest and an auxiliary network to detect occlusive objects in images, automatically determining which images contain occlusive objects without human intervention. This substitution of mechanical manual inspection with automated algorithmic processing directly resolves the contradiction by maintaining high-quality training data selection while eliminating the time cost of manual review.
Solution Approach 2:
The system enables self-service by allowing the data selection process to autonomously evaluate and select images containing occlusive objects. The apparatus automatically processes images through the interest network and auxiliary network, computes reliability scores, and identifies suitable training data without requiring external human labor. This self-service mechanism maintains data quality while dramatically reducing the time investment required for training data curation.
2Reliability
If a large number of images are checked to select occlusive objects, then the completeness of training data selection is improved, but the labor intensity and time consumption increase
Solution Approach 1:
The patent replaces manual image checking with automated computational processing using the interest network and auxiliary network. These networks can process large volumes of images simultaneously and efficiently, computing reliability scores and detecting occlusive objects at speeds far exceeding human capability. This substitution maintains complete coverage of the image dataset while dramatically improving productivity by eliminating the bottleneck of manual inspection.
Solution Approach 2:
The system performs excessive action by automatically evaluating all images in the dataset rather than relying on manual sampling. The automated networks process every image to compute reliability scores and detect occlusive objects, ensuring complete coverage of the data pool. This excessive computational action guarantees completeness of training data selection while the automation maintains high productivity, resolving the contradiction between thoroughness and efficiency.
3Quantity of substance
If artificially created images containing occlusive objects are used, then the availability of training data is improved, but the ability to reflect actual driving conditions is reduced
Solution Approach 1:
The patent applies preliminary action by processing and analyzing actual driving images captured by vehicle sensors before selecting them for training. The interest network and auxiliary network evaluate real images to detect occlusive objects and compute reliability scores, ensuring that the selected training data authentically represents actual driving conditions. This preliminary automated analysis of real-world images maintains both the availability of training data and its representativeness, avoiding the need for artificial image creation.
Data Source
AI summary
Disclosed are an apparatus for collecting training data for image learning and a method thereof. The apparatus may recognize, via an interest network, an object of interest corresponding to a predetermined class by learning an image provided from a vehicle, obtain, via the interest network, one or more reliability scores indicating reliability with which the object of interest is recognized, perform, via an auxiliary network, a learning process associated with the image and detect, in the image, an occlusive object that affects a learning result of the interest network, and determine whether to label the image based on whether the occlusive object is detected and the one or more reliability scores.