Optical flow estimation in the dark with multilevel noise generation

Optical flow estimation in low-light scenes remains a challenging task due to the presence of significant imaging noise, which adversely affects the accuracy and robustness of the estimation. This paper proposes a novel approach for optical flow estimation that leverages Multi-Level Noise Modeling. Initially, the impact of imaging noise on optical flow estimation accuracy is analyzed, leading to the construction of a noisy training dataset specifically tailored for low-light scene optical flow estimation using Multi-Level Noise Modeling techniques. Subsequently, a noise-resistant optical flow estimation network is introduced, designed explicitly for low-light scenarios to improve precision in high-noise environments. The key innovation of this method lies in developing a parameterizable Multi-Level Noise Model and employing implicit feature-supervised training for optical flow estimation under standard lighting conditions, thereby avoiding the need for explicit low-light image enhancement. Experimental results demonstrate that the proposed method exhibits superior noise resistance and robustness across various noise levels, particularly under extreme conditions, where it surpasses existing mainstream methods in foreground-background differentiation and contour edge accuracy.