class ImageSearcher():
'''Image searcher API for clothes retrieval web demo'''
def __init__(self):
root_path = os.path.dirname(__file__)
inds_path = os.path.abspath(os.path.join(root_path, 'db/index'))
feature_path = os.path.abspath(os.path.join(root_path, 'db/feature.npy'))
self.searcher = Searcher(inds_path, feature_path)
self.local_features = np.load('db/local_features.npy')
def search(self, image_path, do_detection=1, k=10):
#queryImage = cv2.imread(image_path)
t1 = Timer()
t1.tic()
#queryFeatures = descriptor.get_descriptor(image_path, multi_box=False)
queryFeatures = descriptor.get_descriptor(image_path)
t1.toc('Feature Extraction time: ')
t2 = Timer()
t2.tic()
#p = Profile()
#results = p.runcall(self.searcher.search, queryFeatures)
#p.print_stats()
results, dists, ind = self.searcher.search(queryFeatures,k=5*k)
#self.reranking(queryFeatures, results, dists, ind, 0.6)
#self.queryExpansion2(results, dists, ind)
#self.queryExpansion(queryFeatures, results, dists, ind, top=3)
t2.toc('Knn search time: ')
result = []
# origine image
#result.append(image_path)
dist = []
for j,imageName in enumerate(results):
if imageName not in result:
result.append(imageName)
dist.append(dists[j])
#print result[:k]
return result[:k],dist[:k]
def reranking(self, queryFeatures, results, dists, ind, rerank_thresh=0.7):
features = self.local_features
feature = []
flag = 0
dist = 0
res = []
for i,index in enumerate(ind):
if dists[i] < rerank_thresh:
flag += 1
else:
if dist == 0:
dist = dists[i-1]
feature.append(features[index])
res.append(results[i])
if len(feature) < 3:
return
feature = np.array(feature).copy()
result,new_ind = self.searcher.research(res, queryFeatures, feature, 3)
for j,r in enumerate(result):
results.insert(flag+j, r)
dists.insert(flag+j, dist)
def queryExpansion2(self, results, dists, ind, threshold=0.3, k=10, top=3):
features = self.searcher.features
feature = []
for i in xrange(top):
query = features[ind[i]]
if dists[i] > threshold:
break
new_result, new_dist, new_ind = self.searcher.search(query,k=k)
for j,dist in enumerate(new_dist):
if dist > threshold:
break
for k,d in enumerate(dists[i:]):
if dist < d:
results.insert(i+k, new_result[j])
dists.insert(i+k, dist)
break
def queryExpansion(self, queryFeatures, results, dists, ind, threshold=0.8, k=10, top=5):
"""
Do Query Expansion with at most top
"""
features = self.searcher.features
feature = []
#feature.append(queryFeatures)
for i,dist in enumerate(dists):
#if dist < threshold and i < top:
if i < top:
feature.append(features[ind[i]])
if len(feature) == 0:
return 0
query = np.mean(np.array(feature), axis=0)
new_results, new_dists, new_ind = self.searcher.search(query,k=k)
for i,dist in enumerate(new_dists):
if dist > dists[-1]:
break
for j,d in enumerate(dists):
if dist < d:
results.insert(j, new_results[i])
dists.insert(j, dist)
break
