Image flow estimation using facet model and covariance propagation

The accuracy of existing methods for gradient-based image how estimation suffers from low-quality derivative estimation, lack of systematic error analysis, and heuristic selection. In this paper we present an image how estimator using the facet model and covariance propagation. The facet model provides high-quality derivatives, image noise variance estimates, and the effect of prefiltering at the same time. We propagate the covariance of the image data to the image flow vector, yielding a covariance matrix with each vector. From the raw flow field, a chi (2) test selects the estimates statistically significant from zero. This selection scheme successfully suppresses false alarms and bad estimates with a low misdetection rate. We further incorporate a flow field smoothness constraint to achieve higher motion field consistency. Experiments on both synthetic and real data, and comparison with other techniques show that at a 10% misdetection rate, our approach has an average error vector magnitude that is 25% less than that of Lucas and Kanade and a false alarm rate less than half of theirs.