From Filters to VLMs: Benchmarking Defogging Methods through Object Detection and Segmentation Performance
A benchmark of nearly 30 defogging methods, measured by their impact on object detection and segmentation in synthetic and real fog.
Ardalan Aryashad, Parsa Razmara, Amin Mahjoub, Seyedarmin Azizi, Mahdi Salmani, Arad Firouzkouhi
5th Workshop on Image, Video, and Audio Quality Assessment · WACV 2026 · Oral
Overview
This benchmark evaluates defogging methods by how much they improve downstream vision tasks such as object detection and segmentation. The study compares performance across synthetic and real-world datasets to reveal gaps in generalization when models trained on synthetic data are applied to real scenes.
Key Findings
The benchmark tests roughly 30 defogging pipelines on Cityscapes (synthetic) and ACDC (real) datasets. Results show that improvements on synthetic data often fail to transfer to real-world conditions, emphasizing the need for evaluation based on downstream task performance rather than visual quality alone.
