We should also remove the concept of "min-num-features", and replace it with a "max-num-features" which sets an upper limit on the number of features. We currently iterate the feature extraction process by lowering the contrast thresholds to achieve a target number of features, but this doesn't necessarily find better features.
DSP SIFT performs better than SIFT and is already implemented in other software like COLMAP. https://arxiv.org/pdf/1412.8556.pdf
It's just waiting to be ported.
We should also remove the concept of "min-num-features", and replace it with a "max-num-features" which sets an upper limit on the number of features. We currently iterate the feature extraction process by lowering the contrast thresholds to achieve a target number of features, but this doesn't necessarily find better features.