def group_feature_extractor(image_list):
"""
Body of the thread that runs the face feature extraction for
a list of images
Arguments:
image_list: List of images to be processed. Each item in the list corresponds to a dictionary with
at least two keys: "path" and "roi". The "path" should contain the full path to the image
file to be processed and "roi" the coordinates of the bounding-box of a face detected on
the image.
"""
list_of_feats = []
if len(image_list) > 0:
try:
# init feature extractor
feature_extractor = face_features.FaceFeatureExtractor()
for image in image_list:
# read image
theim = imutils.acquire_image(image["path"])
det = image["roi"]
# crop image to face detection bounding-box
crop_img = theim[det[1]:det[3], det[0]:det[2], :]
# extract features
feat = feature_extractor.feature_compute(crop_img)
# reshape for compatibility with ranking function
feat = numpy.reshape(feat, (1, settings.FEATURES_VECTOR_SIZE))
# add to list of features to be returned
list_of_feats.append(feat)
except Exception as e:
print 'Exception in group_feature_extractor', e
list_of_feats = []
pass
return list_of_feats
face_detector = face_detection_facenet.FaceDetectorFacenetMTCNN()
# import and create face feature extractor
import face_features
feature_extractor = face_features.FaceFeatureExtractor()
# Compute features for all image paths in args.images_list
all_feats = {'paths': [], 'rois': [], 'feats': []}
with open(args.images_list) as fin:
for img_path in fin:
img_path = img_path.replace('\n', '')
if len(img_path) > 0:
full_path = os.path.join(args.dataset_base_path, img_path)
print 'Computing features for file %s' % (full_path)
# read image
img = imutils.acquire_image(full_path)
# run face detector
detections = face_detector.detect_faces(img)
if numpy.all(detections != None):
for det in detections:
# The coordinates should be already integers, but some basic
# conversion is need for compatibility with all face detectors.
# Plus we have to get rid of the detection score det[4]
det = [int(det[0]), int(det[1]), int(det[2]), int(det[3])]
# crop image to detected face area.
crop_img = img[det[1]:det[3], det[0]:det[2], :]
def addTrs(self, req_params, pos=True):
"""
Adds a training image for the classification process.
Parameters:
req_params: JSON object with at least the fields:
- query_id: the id of the query
- impath: full path to the training image
Other fields include:
- featpath: Full path to the feature file associated to the query
- training_started: boolean indicated that the training step has already started and
therefore the image can be discarded
- extra_params: Another dictionary with the fields:
* from_dataset: Boolean indicating whether the training image
is part of the dataset or not.
* uri: unique resource identifier
* roi: coordinates of a bounding-box defined on the image
the image.
Returns:
JSON formatted string with 'success' field set to 'False'
in case of any problems. The 'success' field set to 'True'
otherwise.
"""
# check query id is present
if 'query_id' in req_params:
query_id = req_params['query_id']
else:
return self.prepare_success_json_str_(False)
# check whether the training step has already started
if self.query_data[str(query_id)]["training_started"]:
# discard the image and return as if nothing happened
if 'impath' in req_params:
print "Training already started. Skipping", os.path.basename(req_params['impath'])
else:
print "Training already started. Skipping image."
return self.prepare_success_json_str_(True)
# check image path is present
if 'impath' in req_params:
impath = req_params['impath']
else:
return self.prepare_success_json_str_(False)
# get the path to the feature file. Not used at the moment.
if 'featpath' in req_params:
featpath = req_params['featpath']
# check for extra parameters
if 'extra_params' in req_params:
if 'from_dataset' in req_params['extra_params']:
from_dataset = req_params['extra_params']['from_dataset']
else:
from_dataset = False
if 'uri' in req_params['extra_params']:
uri = req_params['extra_params']['uri']
else:
uri = -1
if 'roi' in req_params['extra_params']:
# if request specifies a ROI ...
roi = req_params['extra_params']['roi']
roi = numpy.array([int(x) for x in roi]).reshape(len(roi)/2, 2)
xl, yl = roi.min(axis=0)
xu, yu = roi.max(axis=0)
roi = [xl, yl, xu, yu]
print 'Request specifies ROI', roi
# ... check there is a face on the roi
theim = imutils.acquire_image(impath)
crop_img = theim[yl:yu, xl:xu, :]
det = self.face_detector.detect_faces(crop_img, return_best=True)
if numpy.all(det == None):
print 'No detection found in specified ROI'
return self.prepare_success_json_str_(False)
else:
# If found, replace the previous with a more accurate one
det = det[0]
# The coordinates should be already integers, but some basic
# conversion is need for compatibility with all face detectors.
# Plus we have to get rid of the detection score det[4]
det = [int(det[0]), int(det[1]), int(det[2]), int(det[3])]
roi = [det[0]+xl, det[1]+yl, det[2]+xl, det[3]+yl]
print 'Automatically adjusting ROI to more accurate region', roi
else:
roi = None
else:
from_dataset = False
roi = None
uri = -1
# create empty dictionary for the image information
img = dict()
# if the image is brand new ..
if uri == -1:
# and no roi was specified ...
if roi == None:
# read image
theim = imutils.acquire_image(impath)
# run face detector, but only get the best detection.
# multiple detections are not supported for on-the-fly training images
det = self.face_detector.detect_faces(theim, return_best=True)
if numpy.all(det != None):
# if a face is found, save it
print 'Single ROI detected'
# The coordinates should be already integers, but some basic
# conversion is need for compatibility with all face detectors.
# Plus we have to get rid of the detection score det[4]
det = [int(det[0][0]), int(det[0][1]), int(det[0][2]), int(det[0][3])]
print 'final det', det
img["path"] = impath
img["roi"] = det
if pos == True:
img["anno"] = 1
else:
img["anno"] = -1
else:
print 'No detection found'
return self.prepare_success_json_str_(True)
else:
# just the save the image along with the specified roi
img["path"] = impath
img["roi"] = roi
if pos == True:
img["anno"] = 1
else:
img["anno"] = -1
else:
# just the save the image with default values
# for the roi (if not specified) and the annotation type
img["path"] = impath
if roi == None:
img["roi"] = [0, 0, 0, 0]
else:
img["roi"] = roi
if pos == True:
img["anno"] = 1
else:
img["anno"] = -1
# save unique identifier (even if it is -1)
img["uri"] = uri
# save the image information, if we are still accepting training images
if str(query_id) in self.query_data.keys():
if self.query_data[str(query_id)]["training_started"] == False:
self.query_data[str(query_id)]["images"].append(img)
else:
print "Training already started. Skipping", os.path.basename(impath)
else:
print "Query already finished. Skipping", os.path.basename(impath)
# return with success==True
return self.prepare_success_json_str_(True)
