Low-Light Optical Flow Estimation by Enhancing Reflection Component with Global Attention
Zhaolin Xiao, Ke Gui, Kunyu Wang, Haiyan Jin, Haonan Su, Fengyuan Zuo · 2024
This paper presents a novel method for optical flow estimation under low-light conditions, addressing the limitations of existing techniques. Our approach enhances reflection components with global attention within a matching framework, achieving high precision with minimal iterations. We extract global and local motion features by decomposing input images and computing global correlations between adjacent frames, enabling accurate optical flow estimation. A key contribution is the creation of a specialized dataset for lowlight optical flow, featuring synthetic data generated by adjusting brightness and adding noise to standard datasets, alongside real-world low-light videos. This dataset serves as a robust benchmark for evaluating our method. We further optimize computational efficiency using partial convolutions and sparse matching strategies. Experimental results show that our method surpasses existing techniques, particularly in lowbrightness scenarios, validating our approach’s and dataset’s effectiveness.