Data Errors in Major AI Datasets Threaten Computer Vision Reliability, Sejong University Study Finds

A research initiative headed by Sung Wook Baik, a software department professor at Sejong University, has published a comprehensive investigation into labeling flaws embedded within prominent computer vision training benchmarks. The paper, titled "Quality over quantity: a data-centric survey of annotation errors in object detection datasets," appeared in the academic journal Artificial Intelligence Review. Rather than focusing on algorithmic model architecture, the study prioritizes a data-centric approach to evaluate the integrity and reliability of datasets used to train and test object recognition systems.

Object detection serves as a core technical foundation across multiple technology sectors, powering autonomous vehicles, clinical imaging hardware, security camera systems, and industrial robotics. While recent developments in deep learning have dramatically elevated model performance capabilities, the operational dependability of these systems remains closely tied to training data quality. Mislabeled targets or imprecise bounding geometry in benchmark libraries threaten both downstream model deployment and the baseline metrics used to evaluate artificial intelligence algorithms.

The research team systematically evaluated published literature from 2016 through 2025 to catalog methods for uncovering and correcting dataset flaws, as first reported by TechXplore. While earlier academic surveys in the field concentrated heavily on neural network designs, training strategies, or specialized challenges like domain adaptation and few-shot learning, Baik's team focused specifically on identifying data corruption and establishing systematic validation workflows.

To organize the current ecosystem of error detection techniques, the survey classifies existing methodologies into four primary operational groups. Manual strategies rely on direct oversight by human annotators, whereas weakly supervised and semi-supervised routines reduce manual labor by incorporating sparse oversight or unlabeled image banks. Finally, fully automated frameworks utilize model inference outputs, statistical confidence scores, and algorithmic learning mechanisms to detect and rectify dataset mistakes with minimal human involvement.

Beyond synthesizing existing academic literature, lead author Adnan Hussain and the research group conducted a hands-on audit of approximately 47,000 individual images sampled from widely adopted object recognition benchmarks. The empirical evaluation audited several foundational datasets in the artificial intelligence sector, including MS-COCO, DOTA, Open Images, Pascal VOC, FSOD, Objects365, and LVIS.

The audit revealed widespread structural errors across the analyzed benchmark sets. Key issues documented by the team include completely missed labels on clear visual targets, incorrect categorical assignments, and faulty localization where bounding boxes misalign with object edges. The researchers also identified duplicate tags, conflicting labeling conventions, group errors that lump several adjacent objects under a single box, and ambiguous edge cases that prove difficult for human reviewers to standardize.

Such annotation defects corrupt common validation standards by mischaracterizing accurate algorithmic predictions as errors during benchmarking. Furthermore, flawed supervision during training disproportionately hurts a model's ability to recognize small, obscured, rare, or visually complex targets. To resolve these vulnerabilities, the Sejong University team called for new industry standards, including verified benchmark libraries, uniform data quality metrics, context-aware labeling interfaces, cross-dataset harmonization, and automated error remediation assisted by AI foundation models.