Test-time Self Supervision for Data Efficient and Cross-domain Robust Particle Tracking in Turbulent Flow

Abstract

3D particle tracking velocimetry (PTV), which allows the determination of instantaneous velocity, is a key technique for analyzing 3D turbulent flow. While deep learning based methods have fostered advancements in the realm of 3D PTV, they remain heavily dependent on vast volumes of labeled data.This dependency is problematic considering the challenges of collecting particle sets in a large scale, especially when the data collection process is complex or the data itself is rare.Additionally, supervised learning excels primarily with in-domain data, and it may not exhibit the same efficacy with out-of-domain samples.To improve data efficiency and robustness, this paper introduces an innovative test-time self-supervised framework designed for the task of 3D PTV.Notably, we consider an extremely challenging setting of using only 1% data that previous methods use.However, drastically reducing the training samples to this extent can severely hurt performance. To alleviate this, we introduce the zero-divergence loss as an additional regularization mechanism, which draws inspiration from the inherent zero-divergence principle of incompressible fluid's velocity fields.To further enhance cross-domain robustness, we introduce a module named Dynamic Velocimetry Enhancer. During test-time, this module leverages a reconstruction loss to optimize the initial flow predicted by a trained network, enabling the estimated flow to better adjust to unseen samples.Through comprehensive experiments, we reveal that our novel test-time self-supervised framework notably outperforms fully-supervised counterparts with only 1% training samples used by previous methods.Additionally, our cross-domain robustness analysis confirms our framework can naturally generalize to new cases at the merit of test-time optimization. We finally present two evaluations on real-world data in the physical and biological domains, respectively, highlighting the valuable potential for applications in new, noisy real-world scenarios.