Artificial Intelligence Data Quality Research: Errors in Object Detection Datasets
Researchers led by Sejong University examined commonly encountered labeling errors in object detection datasets and solution methods.
A research team led by Professor Sung Wook Baik from the Department of Software at Sejong University has published a comprehensive study examining labeling errors in object detection datasets and methods for detecting and validating these errors.
The Importance of the Data-Centric Approach
The study adopts a data-centric approach focusing on the reliability of datasets used in training and evaluating computer vision models. The performance of these systems, which are used in fields such as autonomous driving, medical imaging, and robotics, is directly dependent on data quality.
Research Scope and Methods
Reviewing related studies published between 2016 and 2025, the researchers categorized existing approaches into four main groups: manual, weakly supervised, semi-supervised, and fully automatic. Additionally, approximately 47,000 images from major benchmark datasets such as MS-COCO and Pascal VOC were examined.
Identified Labeling Errors
As a result of the research, various recurring labeling problems were identified, including missing labels, incorrect labels, localization errors, duplicate labels, and inconsistent labeling.
Future Development Directions
The team outlined future directions such as the development of validated benchmark datasets, standardized quality metrics, and foundation model-assisted error correction systems to improve dataset reliability.