Annotation Verification Using Position Prediction for Image Sequences
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The quality of annotations in image sequences is often compromised by errors or malicious annotators, leading to abnormal annotations that deteriorate the performance of machine learning models, particularly in applications like autonomous driving, where consistent annotation is crucial.
Innovation Solution
An annotation verification method and apparatus that utilize a computer to acquire verified results from adjacent image sequences, predict object positions based on reference information, and compare predicted and actual object positions to verify the accuracy of annotations, thereby identifying and correcting abnormal annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual verification of annotation results is performed for each image in the sequence, then annotation quality can be verified, but the labor required increases significantly
Solution Approach 1:
The system enables automatic self-verification of annotations by using the annotation of one image to predict and verify annotations in adjacent images. The verification apparatus automatically compares predicted object positions with actual annotated positions without requiring manual intervention for each image, thus reducing verification labor while maintaining quality control.
Solution Approach 2:
The system implements a feedback mechanism where the annotation result of a reference image is used to generate predicted annotations for adjacent images, which are then compared with actual annotations to provide verification feedback. This automated feedback loop identifies abnormal annotations without requiring manual verification of every image.
2Quantity of substance
If annotations are performed for all images in a sequence, then comprehensive training data is obtained, but abnormal annotations can significantly deteriorate quality
Solution Approach 1:
The verification apparatus uses feedback from annotated images to automatically detect and identify abnormal annotations in adjacent images. By comparing predicted positions (derived from reference annotations) with actual annotated positions, the system can flag erroneous annotations while maintaining the comprehensive dataset, thus preserving both quantity and quality.
Solution Approach 2:
The system extracts and separates abnormal annotations from the overall annotation dataset by identifying images where the actual annotation deviates significantly from the predicted position. This allows the comprehensive training data to be maintained while extracting and removing or correcting the abnormal annotations that would deteriorate quality.
Data Source
AI summary
The present disclosure is to provide an annotation verification method. In the annotation verification method, a first result which is a verified result of the annotation for a first image sequence included in the image sequence is acquired. Next, first reference information regarding a position of the specified target object range in each image included in the first image sequence is acquired based on the first result. Next, a position of the target object range in a target image is predicted based on reference information including the first reference information. Next, an actually-specified position of the target object range in the target image is acquired based on the result of the annotation for the target image. Then, the result of the annotation for the target image is verified by comparing the target object range at the predicted position and the target object range at the actually-specified position.