Evaluating the impact of dehazing algorithms on object tracking performance

In this study, we present a comprehensive framework to investigate how different image dehazing algorithms impact the performance of visual object tracking methods. We first construct a synthetic hazy video dataset using a 3D simulation engine, which provides paired hazy and haze-free video sequences with moving objects and automatically generated ground-truth annotations. We apply various state-of-the-art dehazing algorithms to these hazy videos and quantitatively evaluate dehazing performance using standard image quality metrics (PSNR and SSIM) against the clear ground-truth frames. Then, we run multiple visual object tracking algorithms on the videos with and without dehazing as a preprocessing step to measure the effect of dehazing on tracking accuracy. Our experiments reveal that while applying dehazing can improve the visual clarity of hazy frames, it does not always correspond to better tracking performance. In some cases, certain dehazing methods even degrade the tracking results for specific trackers. To further validate these findings, we extend our analysis to a real-world video dehazing dataset (REVIDE) where we provide manual object annotations. We observe similar trends on real hazy videos, confirming that the visual quality enhancement and object tracking performance do not have a one-to-one relation.