Robust Fluid Motion Estimator Based on Attentional Transformer
Changdong Yu, Yongpeng Chang, Xiao Liang, Yiwei Fan
Abstract
Estimating complex fluid motions from successive images, i.e., Particle Image Velocimetry (PIV), is of great research significance in physics and engineering applications. Although deep learning-based methods have made great progress for fluid motion estimation, it still remains a challenge in terms of robustness and generalization ability. To address this issue, we present a novel fluid flow estimator called RFMFlowNet in this paper, to robustly estimate complex fluid motion using attentional Transformer. Concretely, the feature encoder based on Transformer is customized to enhance features of two frames globally and stably. Here we also enlarge the receptive field of the encoder according to the characteristics of the flow image to extract more effective information. Then, efficient global matching is performed using 4D correlation volume. Furthermore, a refined flow field is iteratively predicted using a GRU-based optimizer. Extensive experiments, including on synthetic and real-world images, are performed to assess the proposed approach. Experimental results demonstrate that the RFMFlowNet achieves new state-of-the-art performance on the public dataset. Meanwhile, our model enjoys great robustness and generalization ability to challenging flow images, reasonably predicting the evolution trend of fluid flows.