A Systematic Comparison of State-of-the-Art Deep Learning Models for Particle Image Velocimetry

Recent advances in deep learning have led to the development of numerous learning-based Particle Image Velocimetry (PIV) models, which are increasingly gaining attention due to their high spatial resolution, reduced reliance on preprocessing, high accuracy, and robustness. However, a systematic comparison of these models under consistent evaluation conditions remains lacking. In this study, we benchmark several state-of-the-art deep PIV models, including RAFT-PIVs, FlowFormer-PIVs, VideoFlow-PIVs, and LRAFT-STs on a comprehensive dataset of synthetic and experimental flow cases. The evaluation spans a variety of flow types and imaging conditions. Quantitative performance is assessed using averaged-endpoint-error (AEE) against ground truth in synthetic datasets, while some flow characteristics, such as vortex core reconstruction and preservation of turbulent structures, are evaluated compared with traditional cross-correlation PIV algorithm visually. We compare key model architecture and training aspects, such as encoder design, training strategy, and whether the input is pairwise or time-resolved sequences. Our results show that VideoFlow32-PIV, which utilizes an uncompressed encoder, smaller training patches, and multi-frame inputs achieves the lowest overall displacement error. It also excels in capturing small-scale turbulence and maintaining temporal coherence in experimental scenarios. However, we identify several challenging flow cases where all models underperform, highlighting persistent limitations in current approaches. These findings offer practical guidance for selecting appropriate deep PIV models based on specific flow regimes and application requirements, while also identifying promising directions for future development in learning-based PIV methodologies.