For the initital guess of the displacement, at the moment, a FFT based template matching (implemented in scipy) is used, which is the exact same thing that is used in PIV routines and most of the image pair matching in geosciences (COSI-CORR, IMCORR, ...). However, since this only provides a simple initial guess for us, we could define a threshold (e.g. >0.5) for the correlation and perform LSM for all the initial estimates that exhibit a correlation larger than our threshold. In the end, all the LSM results (for each cross-correlation estimate) would need to be compared and the best one selected based on some metric.
Advantages: Probably a more robust working of the method, fewer outliers to begin with, without the need for post-tracking filtering.
Disadvantages: At least some performance losses due to the repeated application of LSM to the same point (depending on the repeatability of large correlation coefficients, this may be significant). It may be the case that this cannot be implemented easily via the "match_template" function from scipy, which would then need a wrapper written by us (or we need to use a different template matching function). This may have large negative impacts on the performance.
For the initital guess of the displacement, at the moment, a FFT based template matching (implemented in scipy) is used, which is the exact same thing that is used in PIV routines and most of the image pair matching in geosciences (COSI-CORR, IMCORR, ...). However, since this only provides a simple initial guess for us, we could define a threshold (e.g. >0.5) for the correlation and perform LSM for all the initial estimates that exhibit a correlation larger than our threshold. In the end, all the LSM results (for each cross-correlation estimate) would need to be compared and the best one selected based on some metric.
Advantages: Probably a more robust working of the method, fewer outliers to begin with, without the need for post-tracking filtering.
Disadvantages: At least some performance losses due to the repeated application of LSM to the same point (depending on the repeatability of large correlation coefficients, this may be significant). It may be the case that this cannot be implemented easily via the "match_template" function from scipy, which would then need a wrapper written by us (or we need to use a different template matching function). This may have large negative impacts on the performance.