Aggregation of patch-based estimations for illumination-invariant optical flow in live cell imaging

Live cell image sequences provide a large variety of challenging situations for motion estimation. We present a novel optical flow estimation method based on a two-stage aggregation framework and designed to handle this diversity of issues. First, semi-local candidates are estimated with a combination of patch correspondences and illumination-invariant affine motion estimations. Then, one candidate is selected at each pixel in a graph-cut based global aggregation stage. This approach allows us to overcome usual limitations of existing methods such as loss of small structures with large displacements, dependency on illumination changes and oversmoothing of the discontinuities. The method jointly estimates the motion and illumination field. We compare our approach to state-of-the-art methods and demonstrate its ability to outperform them in challenging cases frequently arising in live cell imaging.