The retrieval pipeline runs in two passes, each of which can be configured to use any of the three features above: compute features for every image │ first-pass sort (by chosen feature's Euclidean ...