def EMD(saliency_map1, saliency_map2, sub_sample=1/32.0):
'''
Earth Mover's Distance measures the distance between two probability distributions
by how much transformation one distribution would need to undergo to match another
(EMD=0 for identical distributions).
Parameters
----------
saliency_map1 : real-valued matrix
If the two maps are different in shape, saliency_map1 will be resized to match saliency_map2.
saliency_map2 : real-valued matrix
Returns
-------
EMD : float, positive
'''
map2 = np.array(saliency_map2, copy=False)
# Reduce image size for efficiency of calculation
map2 = resize(map2, np.round(np.array(map2.shape)*sub_sample), order=3, mode='nearest')
map1 = resize(saliency_map1, map2.shape, order=3, mode='nearest')
# Histogram match the images so they have the same mass
map1 = match_hist(map1, *exposure.cumulative_distribution(map2))
# Normalize the two maps to sum up to 1,
# so that the score is independent of the starting amount of mass / spread of fixations of the fixation map
map1 = normalize(map1, method='sum')
map2 = normalize(map2, method='sum')
# Compute EMD with OpenCV
# - http://docs.opencv.org/modules/imgproc/doc/histograms.html#emd
# - http://stackoverflow.com/questions/5101004/python-code-for-earth-movers-distance
# - http://stackoverflow.com/questions/12535715/set-type-for-fromarray-in-opencv-for-python
r, c = map2.shape
x, y = np.meshgrid(range(c), range(r))
signature1 = cv.CreateMat(r*c, 3, cv.CV_32FC1)
signature2 = cv.CreateMat(r*c, 3, cv.CV_32FC1)
cv.Convert(cv.fromarray(np.c_[map1.ravel(), x.ravel(), y.ravel()]), signature1)
cv.Convert(cv.fromarray(np.c_[map2.ravel(), x.ravel(), y.ravel()]), signature2)
return cv.CalcEMD2(signature2, signature1, cv.CV_DIST_L2)
def EMD(saliency_map1, saliency_map2, sub_sample=1 / 32.0):
map2 = np.array(saliency_map2, copy=True)
# Reduce image size for efficiency of calculation
map2 = resize(map2,
np.round(np.array(map2.shape) * sub_sample),
order=3,
mode='nearest')
map1 = resize(saliency_map1, map2.shape, order=3, mode='nearest')
# Histogram match the images so they have the same mass
map1 = match_hist(map1, *exposure.cumulative_distribution(map2))
# Normalize the two maps to sum up to 1,
# so that the score is independent of the starting amount of mass / spread of fixations of the fixation map
map1 = normalize(map1, method='sum')
map2 = normalize(map2, method='sum')
# Compute EMD with OpenCV
# - http://docs.opencv.org/modules/imgproc/doc/histograms.html#emd
# - http://stackoverflow.com/questions/5101004/python-code-for-earth-movers-distance
# - http://stackoverflow.com/questions/12535715/set-type-for-fromarray-in-opencv-for-python
r, c = map2.shape
x, y = np.meshgrid(range(c), range(r))
signature1 = cv.CreateMat(r * c, 3, cv.CV_32FC1)
signature2 = cv.CreateMat(r * c, 3, cv.CV_32FC1)
cv.Convert(cv.fromarray(np.c_[map1.ravel(),
x.ravel(),
y.ravel()]), signature1)
cv.Convert(cv.fromarray(np.c_[map2.ravel(),
x.ravel(),
y.ravel()]), signature2)
return cv.CalcEMD2(signature2, signature1, cv.CV_DIST_L2)
def run_wmd_mix(dir):
import numpy as np
import json
k = 10
f = open(str(dir) + "/files.json")
Files = json.loads(f.read())
f2 = open(str(dir) + "/file_cipin_vec_mapper.json")
cipins_mapper = json.loads(f2.read())
f3 = open(str(dir) + "/file_wcd_sim_mapper.json")
wcd_sims_mapper = json.loads(f3.read())
f.close()
f2.close()
f3.close()
del (f)
del (f2)
del (f3)
import gc
gc.collect()
print "Load complete"
from cv2 import cv
sims = []
# import time
# start = time.clock()
from insert_sort import InsertSortItem
count = 0
for file in Files[1800:]:
count += 1
cipin_gram = cipins_mapper[file]
c_list = cipin_gram[0]
d_list = cipin_gram[1]
word_count = len(d_list)
c_recover = np.array(c_list, np.float32).reshape(word_count, -1)
d_recover = np.array(d_list, np.float32)
signature1 = np.column_stack((np.transpose(d_recover), c_recover))
wcd_sims = wcd_sims_mapper[file]
wmd_mix_sims = []
for wcd_sim_gram in wcd_sims[:k]:
file_wcd = wcd_sim_gram[0]
cipin_gram_wcd = cipins_mapper[file_wcd]
c_list_wcd = cipin_gram_wcd[0]
d_list_wcd = cipin_gram_wcd[1]
word_count_wcd = len(d_list_wcd)
c_recover_wcd = np.array(c_list_wcd,
np.float32).reshape(word_count_wcd, -1)
d_recover_wcd = np.array(d_list_wcd, np.float32)
signature2 = np.column_stack(
(np.transpose(d_recover_wcd), c_recover_wcd))
pp = cv.fromarray(signature1)
qq = cv.fromarray(signature2)
emd = cv.CalcEMD2(pp, qq, cv.CV_DIST_L2)
wmd_mix_sims.append([file_wcd, emd])
wmd_mix_sims = sorted(wmd_mix_sims, key=lambda x: x[1], reverse=False)
for wcd_sim_gram in wcd_sims[k:10 * k]:
file_wcd = wcd_sim_gram[0]
cipin_gram_wcd = cipins_mapper[file_wcd]
c_list_wcd = cipin_gram_wcd[0]
d_list_wcd = cipin_gram_wcd[1]
word_count_wcd = len(d_list_wcd)
c_recover_wcd = np.array(c_list_wcd,
np.float32).reshape(word_count_wcd, -1)
d_recover_wcd = np.array(d_list_wcd, np.float32)
# rwmd = getRWMD((c_recover,d_recover),(c_recover_wcd,d_recover_wcd))
# if InsertTest(wmd_mix_sims, rwmd):
signature2 = np.column_stack(
(np.transpose(d_recover_wcd), c_recover_wcd))
pp = cv.fromarray(signature1)
qq = cv.fromarray(signature2)
emd = cv.CalcEMD2(pp, qq, cv.CV_DIST_L2)
InsertSortItem(wmd_mix_sims, [file_wcd, emd])
sims.append([wmd_mix_sims, file])
if count % 100 == 0:
print count * 1.0 / totalFilesNumber / 5 * 2, totalFilesNumber / 5 * 2 - count
# end = time.clock()
# print end-start
f = open(str(dir) + "/wmd_mix_sim_10k.json", "w")
data = json.dumps(sims, ensure_ascii=False)
f.write(data.encode('utf-8'))
def emd_(a32, b32):
return cv.CalcEMD2(a32, b32, cv.CV_DIST_L2)
def EarthMoversDistance(hist1, hist2):
signature1 = _make_signature(hist1)
signature2 = _make_signature(hist2)
return cv.CalcEMD2(signature1, signature2, EUCLIDEAN_DISTANCE)
wcd_sims = wcd_sims_mapper[file]
wmd_mix_sims = [["Nil", 0]]
for wcd_sim_gram in wcd_sims[:k]:
file_wcd = wcd_sim_gram[0]
cipin_gram_wcd = cipins_mapper[file_wcd]
c_list_wcd = cipin_gram_wcd[0]
d_list_wcd = cipin_gram_wcd[1]
word_count_wcd = len(d_list_wcd)
c_recover_wcd = np.array(c_list_wcd,
np.float32).reshape(word_count_wcd, -1)
d_recover_wcd = np.array(d_list_wcd, np.float32)
signature2 = np.column_stack(
(np.transpose(d_recover_wcd), c_recover_wcd))
pp = cv.fromarray(signature1)
qq = cv.fromarray(signature2)
emd = cv.CalcEMD2(pp, qq, cv.CV_DIST_L2)
wmd_mix_sims.append([file_wcd, emd])
top_heap_sort(wmd_mix_sims)
wmd_mix_sims = wmd_mix_sims[1:]
for wcd_sim_gram in wcd_sims[k:10 * k]:
file_wcd = wcd_sim_gram[0]
cipin_gram_wcd = cipins_mapper[file_wcd]
c_list_wcd = cipin_gram_wcd[0]
d_list_wcd = cipin_gram_wcd[1]
word_count_wcd = len(d_list_wcd)
c_recover_wcd = np.array(c_list_wcd,
np.float32).reshape(word_count_wcd, -1)
d_recover_wcd = np.array(d_list_wcd, np.float32)
# rwmd = getRWMD((c_recover,d_recover),(c_recover_wcd,d_recover_wcd))
# if InsertTest(wmd_mix_sims, rwmd):
signature2 = np.column_stack(