A comparative analysis of motion deblurring on human pose estimation performance

Motion blur is a common image degradation that can reduce the accuracy of human pose estimation by weakening body boundaries, edge information, and local joint cues. This study investigates the impact of linear and nonlinear motion blur on human pose estimation, evaluating whether motion deblurring improves keypoint localization performance under different blur severities. We selected a subset of single-person images from the MPII Human Pose Dataset and synthetically applied low, medium, and high levels of linear and nonlinear motion blur to them. Then, we restored the blurred images using three pre-trained deblurring models: RT-Focuser, DeMoE, and NAFNet. We used YOLOv8l-pose as a fixed pose estimator to predict human keypoints on blurred and deblurred images. Since MPII and COCO define keypoints differently, we limited our evaluation to 12 joints that are defined by both datasets. We measured performance using PCKh, mean keypoint Euclidean error (MKEE), an OKS-like score, and joint-wise localization errors. We also assessed the reliability of the deblurring gains using paired image-level bootstrap confidence intervals and Holm-corrected Wilcoxon signed-rank tests. Our results show that pose estimation accuracy decreases as blur severity increases for both linear and nonlinear motion blur. However, the effectiveness of deblurring differs between the two types of blur. Under linear blur, deblurring yields more noticeable improvements at medium and high severity levels. DeMoE yields the most consistent numerical improvements, and RT-Focuser achieves the lowest MKEE under high blur. Under nonlinear blur, DeMoE shows the most consistent numerical improvements, whereas NAFNet and RT-Focuser produce mixed results. Nevertheless, none of the nonlinear blur gains were statistically significant after Holm correction, and all 19 significant comparisons out of 90 occurred under linear blur. Our joint-wise analysis reveals that wrists and ankles are particularly sensitive to stronger motion blur. Overall, our findings suggest that motion deblurring can improve human pose estimation performance; however, its effectiveness depends on the type and severity level of the blur and the selected restoration model.

Authors

Institutions

Publication Details

Journal
Computers and Informatics
Published
2026-11-18
DOI
https://doi.org/10.62189/ci.1988908
Primary Topic
Human Pose and Action Recognition
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A comparative analysis of motion deblurring on human pose estimation performance

Ahmet Selman Bozkır, Mehmet Eren Soykok
5 links
Computers and Informatics
Human Pose and Action Recognition
article

A comparative analysis of motion deblurring on human pose estimation performance

Ahmet Selman Bozkır, Mehmet Eren Soykok
article en

Abstract

Motion blur is a common image degradation that can reduce the accuracy of human pose estimation by weakening body boundaries, edge information, and local joint cues. This study investigates the impact of linear and nonlinear motion blur on human pose estimation, evaluating whether motion deblurring improves keypoint localization performance under different blur severities. We selected a subset of single-person images from the MPII Human Pose Dataset and synthetically applied low, medium, and high levels of linear and nonlinear motion blur to them. Then, we restored the blurred images using three pre-trained deblurring models: RT-Focuser, DeMoE, and NAFNet. We used YOLOv8l-pose as a fixed pose estimator to predict human keypoints on blurred and deblurred images. Since MPII and COCO define keypoints differently, we limited our evaluation to 12 joints that are defined by both datasets. We measured performance using PCKh, mean keypoint Euclidean error (MKEE), an OKS-like score, and joint-wise localization errors. We also assessed the reliability of the deblurring gains using paired image-level bootstrap confidence intervals and Holm-corrected Wilcoxon signed-rank tests. Our results show that pose estimation accuracy decreases as blur severity increases for both linear and nonlinear motion blur. However, the effectiveness of deblurring differs between the two types of blur. Under linear blur, deblurring yields more noticeable improvements at medium and high severity levels. DeMoE yields the most consistent numerical improvements, and RT-Focuser achieves the lowest MKEE under high blur. Under nonlinear blur, DeMoE shows the most consistent numerical improvements, whereas NAFNet and RT-Focuser produce mixed results. Nevertheless, none of the nonlinear blur gains were statistically significant after Holm correction, and all 19 significant comparisons out of 90 occurred under linear blur. Our joint-wise analysis reveals that wrists and ankles are particularly sensitive to stronger motion blur. Overall, our findings suggest that motion deblurring can improve human pose estimation performance; however, its effectiveness depends on the type and severity level of the blur and the selected restoration model.

Computers and Informatics(Advanced Online Publication)
Hacettepe University (TR)
Human Pose and Action Recognition
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