Multi-Criteria Evaluation of Video Deblurring Algorithms for Deployment in Autonomous Robotic Systems
DOI:
https://doi.org/10.64915/RADAP.2026.105.%25pKeywords:
video deblurring, benchmark, autonomous systems, Pareto analysis, temporal stability, ross-dataset validation, roboticsAbstract
This paper presents a methodology for multi-criteria evaluation of video deblurring algorithms for deployment in perception pipelines of autonomous robotic systems, with a primary focus on object detection as the target downstream task.
Motion blur, caused by platform vibration, rapid manoeuvres, and limited illumination, substantially degrades the accuracy of object detectors and other perception modules; a deblurring stage therefore acts as mandatory pre-processing in deployment-critical systems. A systematic comparison of six algorithms – RVRT, Restormer, PVDNet, ESTRNN, DA, and STFAN – is conducted on two datasets (GoPro and DVD) using an extended set of eight metrics that includes traditional restoration quality measures (PSNR, SSIM, LPIPS), temporal stability metrics (tOF ratio, Warping Error), and computational efficiency indicators (FPS, number of parameters).
A Pareto analysis methodology in the space of objectives relevant to the deployment scenario and a weighted utility score with explicitly defined, scenario-dependent weight coefficients are proposed. RVRT achieves the highest frame-level quality on both datasets (30.42 dB on GoPro, 36.08 dB on DVD) and leads on all four quality and temporal stability metrics. For the autonomous detection scenario, PVDNet offers the best quality-throughput-compactness balance on GoPro (29.44 dB at 8.46 FPS, 5.13M parameters), while STFAN joins the Pareto front on DVD (33.11 dB at 8.18 FPS).
Cross-dataset analysis reveals significant changes in the relative ranking of algorithms between datasets: STFAN, the bottom-ranked algorithm on GoPro, gains 5.57 dB on DVD and becomes a Pareto-optimal alternative, while RVRT maintains absolute leadership on both datasets.
The tOF ratio is systematically lower on DVD for all algorithms, indicating more complex temporal dynamics compared to GoPro. Practical recommendations are provided for two deployment scenarios: autonomous object detection, where restoration quality and throughput are prioritised, and operator-in-the-loop control, where per-frame latency is the decisive constraint.
References
1. Zhong Z. et al. (2020). Efficient Spatio-Temporal Recurrent Neural Network for Video Deblurring. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, JM. (eds) Computer Vision – ECCV 2020. ECCV 2020, Lecture Notes in Computer Science, Vol. 12351. Springer, Cham. doi: 10.1007/978-3-030-58539-6_12.
2. Son H. et al. (2021). Recurrent Video Deblurring with Blur-Invariant Motion Estimation and Pixel Volumes. ACM Trans. Graphics (TOG), Vol. 40, Iss. 5, Art. No.: 185, pp. 1–18. doi: 10.1145/3453720.
3. Liang J. et al. (2022). Recurrent Video Restoration Transformer with Guided Deformable Attention. Proc. NeurIPS. doi: 10.48550/arXiv.2206.02146.
4. Zamir S. W. et al. (2022). Restormer: Efficient Transformer for High-Resolution Image Restoration. Proc. CVPR. doi: 10.48550/arXiv.2111.09881.
5. Zhou S. et al. (2019). Spatio-Temporal Filter Adaptive Network for Video Deblurring. Proc. ICCV. doi: 10.48550/arXiv.1904.12257.
6. Kupyn O., Martyniuk T., Wu J., Wang Z. (2019). DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better. Proc. ICCV. DOI: 10.1109/ICCV.2019.00897.
7. Wang Z., Bovik A. C., Sheikh H. R., Simoncelli E. P. (2004). Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Trans. Image Processing, Vol. 13, Iss. 4, pp. 600--612. DOI: 10.1109/TIP.2003.819861.
8. Zhang R., Isola P., Efros A. A., et al. (2018). The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. Proc. CVPR. DOI: 10.1109/CVPR.2018.00068.
9. Sajjadi M. S. M., Vemulapalli R., Brown M. (2018). Frame-Recurrent Video Super-Resolution. Proc. CVPR. DOI: 10.1109/CVPR.2018.00693.
10. Lai W.-S., Huang J.-B., Ahuja N., Yang M.-H. (2016). A Comparative Study for Single Image Blind Deblurring. Proc. CVPR. DOI: 10.1109/CVPR.2016.188.
11. Son H., Lee J., Cho S., Lee S. (2022). Real-Time Video Deblurring via Lightweight Motion Compensation. Computer Graphics Forum (Proc. Pacific Graphics), Vol. 41, Iss. 7, pp. 177-188. doi: 10.1111/cgf.14667.
12. Wu S. et al. (2025). RT-Focuser: A Real-Time Lightweight Model for Edge-side Image Deblurring. Proc. ICTA. DOI: 10.1109/ICTA68203.2025.11329854.
13. Cheng J. et al. (2025). FPGA-Based Real-Time Deblurring and Enhancement for UAV-Captured Infrared Imagery. Remote Sensing, Vol. 17, Iss. 20, 3446. doi: 10.3390/rs17203446.
14. Nah S., Kim T. H., Lee K. M. (2017). Deep Multi-Scale CNN for Dynamic Scene Deblurring. Proc. CVPR. DOI: 10.1109/CVPR.2017.35.
15. Su S., Delbracio M., Wang J., et al. (2017). Deep Video Deblurring for Hand-held Cameras. Proc. CVPR. DOI: 10.1109/CVPR.2017.33.
16. Farneback G. (2003). Two-Frame Motion Estimation Based on Polynomial Expansion. In: Bigun, J., Gustavsson, T. (eds) Image Analysis: Proc. 13th Scandinavian Conference (SCIA), LNCS 2749, Vol. 2749, pp. 363--370, Springer, Berlin, Heidelberg. doi: 10.1007/3-540-45103-X_50.
17. Lai W.-S., Huang J.-B., Wang O., et al. (2018). Learning Blind Video Temporal Consistency. Proc. ECCV. doi: 10.48550/arXiv.1808.00449.
18. Miettinen K. (1999). Nonlinear Multiobjective Optimization. Kluwer Academic Publishers. Def. 2.2.1.
19. Ehrgott M. (2005). Multicriteria Optimization, 2nd ed. Springer. Ch. 2.
20. Boreman G. D. (2001). Modulation Transfer Function in Optical and Electro-Optical Systems. SPIE Press. PDF ISBN: 9780819480453, Print ISBN: 9780819441430. doi: 10.1117/3.419857.
21. Gonzalez R. C., Woods R. E. (2018). Digital Image Processing, 4th ed. Pearson. Ch. 5.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Д. Д. Ляуданський, I. О. Приходько
This work is licensed under a Creative Commons Attribution 4.0 International License.
Authors who publish with this journal agree to the following terms:
- Authors retain copyright and grant the journal right of first publication with the work simultaneously licensed under a Creative Commons Attribution License that allows others to share the work with an acknowledgement of the work's authorship and initial publication in this journal.
- Authors are able to enter into separate, additional contractual arrangements for the non-exclusive distribution of the journal's published version of the work (e.g., post it to an institutional repository or publish it in a book), with an acknowledgement of its initial publication in this journal.
- Authors are permitted and encouraged to post their work online (e.g., in institutional repositories or on their website) prior to and during the submission process, as it can lead to productive exchanges, as well as earlier and greater citation of published work (See The Effect of Open Access).
