kopia lustrzana https://github.com/Shawn-Shan/fawkes
fix minor issue in cropping faces
Shawn-Shan
rodzic
0b663ac422
commit
b9d3ca46da
|
|
@ -4,7 +4,7 @@
|
|||
# @Link : https://www.shawnshan.com/
|
||||
|
||||
|
||||
__version__ = '1.0.1'
|
||||
__version__ = '1.0.2'
|
||||
|
||||
from .differentiator import FawkesMaskGeneration
|
||||
from .protection import main, Fawkes
|
||||
|
|
|
|||
|
|
@ -8,14 +8,13 @@ def to_rgb(img):
|
|||
ret[:, :, 0] = ret[:, :, 1] = ret[:, :, 2] = img
|
||||
return ret
|
||||
|
||||
def aligner():
|
||||
return MTCNN()
|
||||
|
||||
def align(orig_img, aligner, margin=0.8, detect_multiple_faces=True):
|
||||
def aligner():
|
||||
return MTCNN(min_face_size=30)
|
||||
|
||||
|
||||
def align(orig_img, aligner):
|
||||
""" run MTCNN face detector """
|
||||
minsize = 20 # minimum size of face
|
||||
threshold = [0.6, 0.7, 0.7] # three steps's threshold
|
||||
factor = 0.709 # scale factor
|
||||
|
||||
if orig_img.ndim < 2:
|
||||
return None
|
||||
|
|
@ -23,45 +22,57 @@ def align(orig_img, aligner, margin=0.8, detect_multiple_faces=True):
|
|||
orig_img = to_rgb(orig_img)
|
||||
orig_img = orig_img[:, :, 0:3]
|
||||
|
||||
bounding_boxes = aligner.detect_faces(orig_img)
|
||||
nrof_faces= len(bounding_boxes)
|
||||
detect_results = aligner.detect_faces(orig_img)
|
||||
cropped_arr = []
|
||||
bounding_boxes_arr = []
|
||||
for dic in detect_results:
|
||||
if dic['confidence'] < 0.9:
|
||||
continue
|
||||
x, y, width, height = dic['box']
|
||||
if width < 30 or height < 30:
|
||||
continue
|
||||
bb = [y, x, y + height, x + width]
|
||||
cropped = orig_img[bb[0]:bb[2], bb[1]:bb[3], :]
|
||||
cropped_arr.append(np.copy(cropped))
|
||||
bounding_boxes_arr.append(bb)
|
||||
return cropped_arr, bounding_boxes_arr
|
||||
|
||||
if nrof_faces > 0:
|
||||
det = bounding_boxes[0]['box']
|
||||
det_arr = []
|
||||
img_size = np.asarray(orig_img.shape)[0:2]
|
||||
if nrof_faces > 1:
|
||||
margin = margin / 1.5
|
||||
if detect_multiple_faces:
|
||||
for i in range(nrof_faces):
|
||||
det_arr.append(np.squeeze(bounding_boxes[i]['box']))
|
||||
else:
|
||||
bounding_box_size = (det[1] + det[3])
|
||||
img_center = img_size / 2
|
||||
offsets = np.vstack([(det[0] + det[2]) / 2 - img_center[1],
|
||||
(det[1] + det[3]) / 2 - img_center[0]])
|
||||
offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
|
||||
index = np.argmax(bounding_box_size - offset_dist_squared * 2.0) # some extra weight on the centering
|
||||
det_arr.append(det[index, :])
|
||||
else:
|
||||
det_arr.append(np.squeeze(det))
|
||||
|
||||
cropped_arr = []
|
||||
bounding_boxes_arr = []
|
||||
for i, det in enumerate(det_arr):
|
||||
det = np.squeeze(det)
|
||||
bb = np.zeros(4, dtype=np.int32)
|
||||
# add in margin
|
||||
marg1 = int((det[2] - det[0]) * margin)
|
||||
marg2 = int((det[3] - det[1]) * margin)
|
||||
|
||||
bb[0] = max(det[0] - marg1/2, 0)
|
||||
bb[1] = max(det[1] - marg2/2, 0)
|
||||
bb[2] = min(det[0] + det[2] + marg1/2, img_size[0])
|
||||
bb[3] = min(det[1] + det[3] + marg2/2, img_size[1])
|
||||
cropped = orig_img[bb[0]:bb[2], bb[1]: bb[3],:]
|
||||
cropped_arr.append(cropped)
|
||||
bounding_boxes_arr.append([bb[0], bb[1], bb[2], bb[3]])
|
||||
return cropped_arr, bounding_boxes_arr
|
||||
else:
|
||||
return None
|
||||
# if nrof_faces > 0:
|
||||
# det = bounding_boxes[0]['box']
|
||||
# det_arr = []
|
||||
# img_size = np.asarray(orig_img.shape)[0:2]
|
||||
# if nrof_faces > 1:
|
||||
# margin = margin / 1.5
|
||||
# if detect_multiple_faces:
|
||||
# for i in range(nrof_faces):
|
||||
# det_arr.append(np.squeeze(bounding_boxes[i]['box']))
|
||||
# else:
|
||||
# bounding_box_size = (det[1] + det[3])
|
||||
# img_center = img_size / 2
|
||||
# offsets = np.vstack([(det[0] + det[2]) / 2 - img_center[1],
|
||||
# (det[1] + det[3]) / 2 - img_center[0]])
|
||||
# offset_dist_squared = np.sum(np.power(offsets, 2.0), 0)
|
||||
# index = np.argmax(bounding_box_size - offset_dist_squared * 2.0) # some extra weight on the centering
|
||||
# det_arr.append(det[index, :])
|
||||
# else:
|
||||
# det_arr.append(np.squeeze(det))
|
||||
#
|
||||
# cropped_arr = []
|
||||
# bounding_boxes_arr = []
|
||||
# for i, det in enumerate(det_arr):
|
||||
# det = np.squeeze(det)
|
||||
# bb = np.zeros(4, dtype=np.int32)
|
||||
# # add in margin
|
||||
# marg1 = int((det[2] - det[0]) * margin)
|
||||
# marg2 = int((det[3] - det[1]) * margin)
|
||||
#
|
||||
# bb[0] = max(det[0] - marg1 / 2, 0)
|
||||
# bb[1] = max(det[1] - marg2 / 2, 0)
|
||||
# bb[2] = min(det[0] + det[2] + marg1 / 2, img_size[0])
|
||||
# bb[3] = min(det[1] + det[3] + marg2 / 2, img_size[1])
|
||||
# cropped = orig_img[bb[0]:bb[2], bb[1]: bb[3], :]
|
||||
# cropped_arr.append(cropped)
|
||||
# bounding_boxes_arr.append([bb[0], bb[1], bb[2], bb[3]])
|
||||
# return cropped_arr, bounding_boxes_arr
|
||||
# else:
|
||||
# return None
|
||||
|
|
|
|||
|
|
@ -57,14 +57,14 @@ class Fawkes(object):
|
|||
|
||||
def mode2param(self, mode):
|
||||
if mode == 'low':
|
||||
th = 0.0045
|
||||
max_step = 25
|
||||
th = 0.005
|
||||
max_step = 30
|
||||
lr = 25
|
||||
extractors = ["extractor_2"]
|
||||
|
||||
elif mode == 'mid':
|
||||
th = 0.012
|
||||
max_step = 50
|
||||
max_step = 75
|
||||
lr = 20
|
||||
extractors = ["extractor_0", "extractor_2"]
|
||||
|
||||
|
|
|
|||
|
|
@ -145,13 +145,8 @@ class Faces(object):
|
|||
p = image_paths[i]
|
||||
self.org_faces.append(cur_img)
|
||||
|
||||
if eval_local:
|
||||
margin = 0
|
||||
else:
|
||||
margin = 0.7
|
||||
|
||||
if not no_align:
|
||||
align_img = align(cur_img, self.aligner, margin=margin)
|
||||
align_img = align(cur_img, self.aligner)
|
||||
if align_img is None:
|
||||
print("Find 0 face(s) in {}".format(p.split("/")[-1]))
|
||||
self.images_without_face.append(i)
|
||||
|
|
@ -221,6 +216,7 @@ class Faces(object):
|
|||
org_shape = self.cropped_faces_shape[i]
|
||||
|
||||
old_square_shape = max([org_shape[0], org_shape[1]])
|
||||
|
||||
cur_protected = resize(cur_protected, (old_square_shape, old_square_shape))
|
||||
cur_original = resize(cur_original, (old_square_shape, old_square_shape))
|
||||
|
||||
|
|
|
|||
Ładowanie…
Reference in New Issue