Low-Light Optical Flow Estimation Method Based on Implicit Modeling of Mixed Noises

This paper presents a data-driven solution to address the challenge of decreased accuracy in optical flow estimation under low-light conditions. It proposes synthesizing a dataset with multiple mixed noises to train a neural network for optical flow estimation. The neural network includes a noise removal module to enhance its performance. Experimental results demonstrate the superiority of this approach, particularly in datasets with higher levels of noise. By incorporating a noise removal module into the optical flow estimation network, the proposed method effectively mitigates the impact of noise on accuracy. Additionally, the synthesis of a dataset with mixed noises allows for robust training of the neural network, improving its performance in real-world low-light scenarios.