Impact of Bounding Box Annotation Noise on Object Detection Performance: A Multi-Seed Evaluation of YOLO11n on a COCO2017 Subset

This study investigates the impact of controlled bounding-box annotation noise on YOLO11n object detection performance using a COCO2017-derived subset. The experimental dataset comprises 1,500 images, including 1,000 training images, 200 validation images, and 300 independent test images. Four training annotation noise conditions (0%, 5%, 10%, and 20%) were evaluated across five matched random seeds (42, 123, 456, 789, and 2026), resulting in 20 training experiments. Controlled perturbations were applied to training bounding-box annotations, while validation and independent test annotations remained unchanged. The independent test-set mean mAP50–95 scores were 0.395969, 0.396229, 0.395525, and 0.394158 for the 0%, 5%, 10%, and 20% noise conditions, respectively. Paired statistical comparisons, with Holm correction applied to the three primary comparisons against the clean baseline, did not identify statistically significant differences. The observed performance differences were small within the evaluated experimental setting. However, the results do not demonstrate statistical equivalence or establish general robustness to annotation noise. The findings are limited to the selected dataset subset, model architecture, noise mechanism, training configuration, and five random seeds. This independent research provides an exploratory evaluation of bounding-box annotation quality in object detection and discusses statistical uncertainty, experimental limitations, and reproducibility considerations. Research repository: https://github.com/tanvirshadgit/bounding-box-noise-research Research status: Independent research preprint; not peer-reviewed.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24224385
Primary Topic
Advanced Neural Network Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Impact of Bounding Box Annotation Noise on Object Detection Performance: A Multi-Seed Evaluation of YOLO11n on a COCO2017 Subset

Tanvir Ahmed Khan
Zenodo (CERN European Organization for Nuclear Research)
Advanced Neural Network Applications
preprint

Impact of Bounding Box Annotation Noise on Object Detection Performance: A Multi-Seed Evaluation of YOLO11n on a COCO2017 Subset

Tanvir Ahmed Khan
preprint en

Abstract

This study investigates the impact of controlled bounding-box annotation noise on YOLO11n object detection performance using a COCO2017-derived subset. The experimental dataset comprises 1,500 images, including 1,000 training images, 200 validation images, and 300 independent test images. Four training annotation noise conditions (0%, 5%, 10%, and 20%) were evaluated across five matched random seeds (42, 123, 456, 789, and 2026), resulting in 20 training experiments. Controlled perturbations were applied to training bounding-box annotations, while validation and independent test annotations remained unchanged. The independent test-set mean mAP50–95 scores were 0.395969, 0.396229, 0.395525, and 0.394158 for the 0%, 5%, 10%, and 20% noise conditions, respectively. Paired statistical comparisons, with Holm correction applied to the three primary comparisons against the clean baseline, did not identify statistically significant differences. The observed performance differences were small within the evaluated experimental setting. However, the results do not demonstrate statistical equivalence or establish general robustness to annotation noise. The findings are limited to the selected dataset subset, model architecture, noise mechanism, training configuration, and five random seeds. This independent research provides an exploratory evaluation of bounding-box annotation quality in object detection and discusses statistical uncertainty, experimental limitations, and reproducibility considerations. Research repository: https://github.com/tanvirshadgit/bounding-box-noise-research Research status: Independent research preprint; not peer-reviewed.

Zenodo (CERN European Organization for Nuclear Research)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon