Rough LSH python benchmark of potential ORB descriptors

lsh-benchmark-test.py

from lshash import LSHash
import numpy
import time

lsh = LSHash(100, 200, 1, None, "hamming.npz")

# Very modest
# lets assume 5000 images, with 200 features and 200 binary descriptors each
for y in range(0, 5000):
  for x in range(0, 200):
    lsh.index(numpy.random.randint(2, size=(200,)), y)

for r in range(0,10):
  now = time.time()
  lsh.query(numpy.random.randint(2, size=(200,)), 10, "hamming")
  print time.time()-now

  
# 1.9097840786
# 2.78951191902
# 0.34514093399
# 1.41738390923
# 2.75403499603
# 1.3472340107
# 3.9218878746
# 0.842194795609
# 1.44249796867
# 1.45863008499

20,000 image sets (200*200 binary feature per image)

6.6Gb RAM usage

10285 root      20   0 7289292 6.607g   5616 R  99.7 42.5   2:03.41 python   

timing in Seconds per single query (200* would be required, one for each feature?)

16.2959368229
1.22037315369
1.19355297089
4.36228394508
5.90254998207
1.46862792969
17.2071578503
4.79945707321
6.87375688553
16.2232999802