Challenges of deep unsupervised optical flow estimation for particle-image velocimetry data

In recent years, several algorithms have been proposed that leverage deep learning techniques within the analysis workflow of particle-image velocimetry (PIV) measurements. This emerging body of work has shown that deep learning has the potential to match or outperform state-of-the-art classical algorithms in terms of efficiency, accuracy, and spatial resolution. However, the huge diversity in dynamic flows and varying particle-image conditions require PIV processing schemes to have high generalization capabilities to unseen flow and lighting conditions. If these conditions vary strongly compared to the training data, the performance of fully supervised PIV tools can degrade substantially. In contrast, unsupervised learning ameliorates the need for comprehensive labeled training data and can permit a much wider range of data to be used during training. Therefore, unsupervised deep learning could improve inference capability for challenging real-world use cases. However, design of an unsupervised loss objective is non-trivial and requires application-specific consideration. Motivated by the foregoing, in this paper we study unsupervised deep learning for PIV processing, systematically investigating key components of losses and accommodating regularizers and deriving a proxy loss. The resulting algorithm, named Unsupervised Recurrent All-Pairs Field Transforms for PIV (URAFT-PIV), is unsupervised and meant specifically for PIV applications. We investigate performance under varying image and lighting conditions in synthetic and experimental data, with a breadth and depth going well beyond currently available empirical results. These results shed new light on deep learning for PIV processing and in particular on the scope for unsupervised learning in this domain.