def Classifier(self,
database,
version,
num_of_kpts,
cutoff,
top = 5,
K = 10,
depth=4):
num_in_set = database[0]
num_of_sets = database[1]
test_database = database[3]
total = num_in_set * num_of_sets
num_in_test_set = len(test_database) / num_of_sets
classify_score =np.zeros(num_of_sets)
class_name = [[] for i in range(num_of_sets)]
#load weight
wt = Weight(cutoff)
sign_dir = os.path.join(self.SIGN_DIR, 'db_version_' + str(version))
wt.load_weights(sign_dir, self.WEIGHT_FILE+str(num_of_kpts)+'_'+str(cutoff))
wt.load_weighted_sign(sign_dir, self.WEIGHT_SIGN_FILE+str(num_of_kpts)+'_'+str(cutoff))
#Load tree
tree_dir = os.path.join(self.TREE_DIR, 'db_version_' + str(version))
tr = vl._vlfeat.VlHIKMTree(0, 0)
tr.load(os.path.join(tree_dir, str(num_of_kpts) + self.TREE_FILE))
for k in test_database:
#randomly get image from the img_dir
img = Image.open(k[2]).convert('L')
img = self.StandalizeImage(img, 480)
img_data = np.asarray(img, dtype=float)
#generate desc, sign and weighted sign
kp = Keypoint()
#kp.load_keypoint(self.KEYPOINT_DIR, self.KEYPOINT_FILE+str(num_of_kpts))
kp.generate_keypoint(num_of_kpts, img.size[0], img.size[1], self.SIGMA)
desc = Descriptor()
desc.generate_desc(img_data, kp.kpt)
#very important !! convert desc to float type
#desc_f = np.array(desc.desc, dtype=float)
sign = Signature()
s = sign.generate_sign(tr,desc.desc, K, depth)
weighted_sign = wt.weight_sign(s)
#vote
d=np.empty(total)
for i in range(0, total):
d[i] = self.dist(wt.weighted_sign[i,:], weighted_sign)
perm = np.argsort(d)
vote_for = np.floor((perm[0:top])/num_in_set)+1
votes = vl.vl_binsum(np.zeros(num_of_sets), np.ones(top), vote_for)
#print votes
best = np.argmax(votes)
if best == k[1]:
classify_score[k[1]] += 1
print '=>'+str(k[0])
class_name[k[1]] = k[3]
classify_score = classify_score / num_in_test_set
return zip(class_name, classify_score.tolist())
