Iterative Filtering and Smoothing with Optical Flow Prediction Models

Our paper, “Iterative Filtering and Smoothing with Optical Flow Prediction Models,” has been published in SIAM Journal on Imaging Sciences.

In this paper, we propose a new data assimilation approach based on iterative filtering and smoothing in the expectation-maximization fashion, incorporating optical flow prediction models for dynamic state estimation. The concept is suitable for applications where state estimation relies on 2D images and where no natural physical prediction model is available. We apply the proposed approach to dynamic X-ray images and both real and synthetic satellite data, demonstrating that iterative filtering and smoothing with optical flow models yields improved results compared to using an identity model approach. The quality of the state estimates improves both visually and in terms of the root mean squared error, typically after a couple of iterations. However, in some cases, continued iterations may lead to deteriorating results. Therefore, monitoring the quality of both optical flow and state estimates is crucial in the iterative approach.

Citation:
Janne Hakkarainen, Zenith Purisha, Neus Sabater, Monika Szeląg, Samuli Siltanen and Antti Solonen: Iterative Filtering and Smoothing with Optical Flow Prediction Models, SIAM Journal on Imaging Sciences, Volume 18, Number 4, doi:10.1137/24M1689120, 2025. https://epubs.siam.org/doi/full/10.1137/24M1689120