Jowekk

I also encountered this problem, and I felt that hnsw's recall under the large data set was seriously reduced. I have 5000 pairs of face matching features. I mix one...

I checked this discussion [https://github.com/nmslib/hnswlib/issues/164](url). As to whether there are duplicate data, I use 5 million gallery data sets to retrieve themselves: top 1: 4861229/5000000 = 0.972246 top 5: 4999996/5000000...

5000 pairs: paired images taken by 5000 different people. gallery data: completely unrelated to the above 5000 people. I used 5000 images as batch queries to recall the other 5000...

ef = 10000 is too big to meet our time requirements. When ef and gallery size are increased, a large number of random accesses will be caused because neighbors are...

ok, this is a good idea worth trying. thank u