更好的AI始于更好的数据：研究人员识别目标检测数据集中的隐藏标注错误

September 7, 2026 Better AI starts with better data: Researchers identify hidden annotation errors in object detection datasets by Sejong University edited by Sadie Harley, reviewed by Robert Egan Sadie Harley Scientific Editor Meet our editorial team Behind our editorial process Robert Egan Senior Editor Meet our editorial team Behind our editorial process Editors' notes This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: fact-checked trusted source proofread The GIST Add as preferred source An illustration of sample images from each dataset, highlighting their diversity. Credit: Artificial Intelligence Review (2026). DOI: 10.1007/s10462-026-11502-z A research team led by Sung Wook Baik, a professor in the Department of Software at Sejong University, has published a comprehensive survey examining annotation errors in widely used object detection datasets and methods for identifying and validating them. The paper, titled "Quality over quantity: a data-centric survey of annotation errors in object detection datasets," was published in Artificial Intelligence Review. The study takes a data-centric approach to object detection by focusing on the quality and reliability of the datasets used to train and evaluate computer vision models. Object detection is a fundamental computer vision technology used in autonomous driving, medical imaging, surveillance, robotics and many other applications. Although advances in deep learning have significantly improved detection performance, the reliability of these systems depends heavily on the quality of their training data. Even widely used benchmark datasets can contain inaccurate or incomplete annotations that may affect both model training and performance evaluation. Previous object detection surveys have primarily focused on model architecture, training strategies, evaluation metrics and challenges such as few-shot learning, domain adaptation and open-world detection. In contrast, Baik and his team focused specifically on the detection and validation of annotation errors, reviewing related studies published between 2016 and 2025. Evolution of popular object detection and image annotation datasets from 2005 to 2025. The timeline illustrates key milestones, including the release of Pascal VOC, MS COCO, Open Images, FSOD, and emerging datasets such as Sama-COCO and MJ-COCO. Credit: Artificial Intelligence Review (2026). DOI: 10.1007/s10462-026-11502-z The researchers categorized existing approaches into four major groups: manual, weakly supervised, semi-supervised and fully automatic methods. Manual approaches rely on direct human inspection, while weakly- and semi-supervised techniques reduce annotation requirements through limited supervision or unlabeled data. Fully automatic methods use model predictions, confidence estimates and other learning-based mechanisms to detect or correct annotation problems with minimal human intervention. Beyond reviewing the literature, the team also manually examined around 47,000 images from major object detection benchmarks. Their analysis included datasets such as MS-COCO, DOTA, Open Images, Pascal VOC, FSOD, Objects365 and LVIS. The study identified several recurring types of annotation problems. These include missed annotations, where visible objects are not labeled; incorrect labels, where objects are assigned to the wrong category; and localization errors, where bounding boxes do not correctly align with objects. The researchers also observed duplicate annotations, inconsistent annotations, group errors, where multiple nearby objects are covered by a single bounding box, and debatable examples involving visually ambiguous or difficult cases. Such errors can affect commonly used evaluation measures and may cause correct model predictions to be treated as errors. Inaccurate labels and bounding boxes can also introduce unreliable supervision during training, with particularly significant effects on small, occluded, rare or visually ambiguous objects. The researchers also outlined several future directions for improving dataset reliability. These include developing verified benchmark datasets, standardized annotation-quality metrics, context-aware annotation tools, active human-model feedback systems, cross-dataset annotation harmonization and foundation-model-assisted error correction. The study highlights the need for clearer standards for bounding boxes, label hierarchies and visibility thresholds for partially visible objects. The survey is expected to provide researchers and practitioners with a systematic reference for improving annotation quality and developing more reliable, robust and scalable object detection systems. More information Adnan Hussain et al, Quality over quantity: a data-centric survey of annotation errors in object detection datasets, Artificial Intelligence Review (2026). DOI: 10.1007/s10462-026-11502-z Key concepts Large language modelsComputational 3D visionTrustworthy machine learningMachine learning methodologies Provided by Sejong University Who's behind this story? Sadie Harley BSc Life Sciences & Ecology. Microbiology lab background with pharmaceutical news experience in oil, gas, and renewable industries. Full profile → Robert Egan Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile → Citation: Better AI starts with better data: Researchers identify hidden annotation errors in object detection datasets (2026, September 7) retrieved 8 September 2026 from https://techxplore.com/news/2026-09-ai-hidden-annotation-errors-datasets.html This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only. Automatic label checking: The missing step in making medical AI reliable Dec 24, 2025 AI models for remote object detection are vulnerable to both physical and digital attacks, research finds Oct 17, 2024 New method allows robot vision to identify occluded objects Apr 19, 2022 Comparing AI anatomy segmentation models when ground truth is missing May 13, 2026 Squashing 'fantastic bugs' hidden in AI benchmarks Dec 11, 2025 New remote sensing dataset improves global land change tracking Nov 22, 2023 Memristor chip breaks the capacity limit of brain-inspired associative memory 6 hours ago Invention may bring smart bionic eye into sight 5 hours ago Whole-body expansion and contraction make a robot seem more alive 7 hours ago A tiny crystal tilt may bring greener magnetic memory closer to everyday devices 7 hours ago Tracking energy flow reveals voltage pattern that slashes waste heat by over 60% 9 hours ago Editorial proposes limits on generative AI's role in scientific research 13 hours ago