Generalization of deep recurrent optical flow estimation for particle-image velocimetry data

Abstract Particle-image velocimetry (PIV) is one of the key techniques in modern experimental fluid mechanics to determine the velocity components of flow fields in a wide range of complex engineering problems. Current PIV processing tools are mainly handcrafted models based on cross-correlations computed across interrogation windows. Although widely used, these existing tools have a number of well-known shortcomings, including limited spatial output resolution and peak-locking biases. Recently, new approaches for PIV processing leveraging a novel neural network architecture for optical flow estimation called recurrent all-pairs field transforms (RAFT) have been developed. These have matched or exceeded the performance of classical, handcrafted models. While the RAFT-PIV method is a promising approach, it is important for the broader fluids community to more completely understand its empirical behavior and performance. To this end, in this study, we thoroughly investigate the performance of RAFT-PIV under varying image and lighting conditions. We consider applications spanning synthetic and experimental data, with a breadth and depth going far beyond currently available empirical results. The results for the wide variation of experiments shed new light on the capabilities of deep learning for PIV processing.