def extract_region(self, image):
DIMS = image.shape
blob = cv2.dnn.blobFromImage(
image,
1.0,
(300, 300) # , [104, 117, 123], False, False
)
self.net.setInput(blob)
detections = self.net.forward()[0, 0, :, :]
regions = []
for i in range(detections.shape[0]):
if detections[i, 2] >= self.threshold:
reg = region.RectangleRegion(
detections[i, 3] * DIMS[1] - self.margin / 2,
detections[i, 4] * DIMS[0] - self.margin / 2,
(detections[i, 5] - detections[i, 3]) * DIMS[1] +
self.margin,
(detections[i, 6] - detections[i, 4]) * DIMS[0] +
self.margin,
detections[i, 2],
)
logging.debug("Face {} detected: {}".format(i, str(reg)))
regions.append(reg)
return regions, detections
def extract_region(
self,
image,
threshold=0.9, # image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
):
bbs, points = mtcnn_tensorflow.detect_face(
image,
self.min_size,
self.pnet,
self.rnet,
self.onet,
self.threshold,
self.factor,
)
regions = [
region.RectangleRegion(
bb[0], bb[1], bb[2] - bb[0], bb[3] - bb[1], bb[4]
)
for bb in bbs
if bb[4] >= threshold
]
return regions, (bbs, points)
def predict(self,
X: np.ndarray,
threshold=0.95,
extract_image=False) -> datum.Datum:
ret = self.detector.detect_face(X)
if ret is None:
return []
bbox, points = ret
if bbox.shape[0] == 0:
return []
data = []
for bb, point in zip(*ret):
if bb[4] >= threshold:
dat = datum.Datum(
region=region.RectangleRegion(bb[0], bb[1], bb[2] - bb[0],
bb[3] - bb[1], bb[4]),
keypoints=point,
)
if extract_image:
point = point.reshape((2, 5)).T
aligned = align_mtcnn.align(X,
bb,
point,
image_size="112,112")
dat.image = aligned
data.append(dat)
return data
def extract_region(self, image):
raw_detection = self.face_det(image, 0)
logger.info("Number of faces detected: {}".format(len(raw_detection)))
regions = [
region.RectangleRegion(
i.rect.left(),
i.rect.top(),
i.rect.right() - i.rect.left(),
i.rect.bottom() - i.rect.top(),
) for i in raw_detection
]
return regions, raw_detection
def update(self, image):
res = []
rem_list = []
for i, tr in enumerate(self.tracklets):
retval, bbox = tr.tracker.update(image)
if retval:
reg = region.RectangleRegion(
bbox[0], bbox[1], bbox[2], bbox[3]
)
res.append(reg)
else:
# DONE Destroy tracker condition
logger.debug("Tracklet face retrieval faillure.")
tr.ttd -= 1
if tr.ttd == 0:
rem_list.append(i)
# Remove all tracklets where ttd == 0
for i in rem_list:
self.tracklets.pop(i)
return res
def parse_region_line(line, type=2):
contents = line[:-1].split()
if type == 1:
# Rectangle
dat = region.RectangleRegion(
float(contents[0]),
float(contents[1]),
float(contents[2]),
float(contents[3]),
float(contents[4]),
)
elif type == 2:
# Ellipse
dat = region.EllipseRegion(
float(contents[0]),
float(contents[1]),
float(contents[2]),
float(contents[3]),
float(contents[4]),
float(contents[5]),
)
else:
raise ValueError()
return dat
def openpose_face_detector(posePtr, threshold):
point_top_left = np.zeros(2)
face_size = 0.0
score = 0.0
points_used = set()
neckScoreAbove = posePtr[1, 2] > threshold
headNoseScoreAbove = posePtr[0, 2] > threshold
lEarScoreAbove = posePtr[16, 2] > threshold
rEarScoreAbove = posePtr[17, 2] > threshold
lEyeScoreAbove = posePtr[14, 2] > threshold
rEyeScoreAbove = posePtr[15, 2] > threshold
counter = 0.0
if neckScoreAbove and headNoseScoreAbove:
if (lEyeScoreAbove == lEarScoreAbove
and rEyeScoreAbove == rEarScoreAbove
and lEyeScoreAbove != rEyeScoreAbove):
if lEyeScoreAbove:
point_top_left += (posePtr[14, 0:2] + posePtr[16, 0:2] +
posePtr[0, 0:2]) / 3.0
face_size += 0.85 * (dist(posePtr[14, 0:2], posePtr[16, 0:2]) +
dist(posePtr[0, 0:2], posePtr[16, 0:2]) +
dist(posePtr[14, 0:2], posePtr[0, 0:2]))
points_used = points_used.union([0, 14, 16])
else:
point_top_left += (posePtr[15, 0:2] + posePtr[17, 0:2] +
posePtr[0, 0:2]) / 3.0
face_size += 0.85 * (dist(posePtr[15, 0:2], posePtr[17, 0:2]) +
dist(posePtr[0, 0:2], posePtr[17, 0:2]) +
dist(posePtr[15, 0:2], posePtr[0, 0:2]))
points_used = points_used.union([0, 15, 17])
else:
point_top_left += (posePtr[1, 0:2] + posePtr[0, 0:2]) / 2.0
face_size += 2.0 * dist(posePtr[1, 0:2], posePtr[0, 0:2])
points_used = points_used.union([0, 1])
counter += 1.0
if lEyeScoreAbove and rEyeScoreAbove:
point_top_left += (posePtr[14, 0:2] + posePtr[15, 0:2]) / 2.0
face_size += 3.0 * dist(posePtr[14, 0:2], posePtr[15, 0:2])
counter += 1.0
points_used = points_used.union([14, 15])
if lEarScoreAbove and rEarScoreAbove:
point_top_left += (posePtr[16, 0:2] + posePtr[17, 0:2]) / 2.0
face_size += 2.0 * dist(posePtr[16, 0:2], posePtr[17, 0:2])
counter += 1.0
points_used = points_used.union([16, 17])
if counter > 0:
point_top_left /= counter
face_size /= counter
score = np.mean(posePtr[list(points_used), 2])
return region.RectangleRegion(
point_top_left[0] - face_size / 2,
point_top_left[1] - face_size / 2,
face_size,
face_size,
score,
)
def openpose_face_detector2(posePtr, threshold):
point_top_left = np.zeros(2)
face_size = [0.0, 0.0] # (width, height)
score = np.mean(posePtr[[0, 1, 14, 15, 16, 17], 2])
l_true = None
r_true = None
neckScoreAbove = posePtr[1, 2] > threshold
headNoseScoreAbove = posePtr[0, 2] > threshold
lEarScoreAbove = posePtr[16, 2] > threshold
rEarScoreAbove = posePtr[17, 2] > threshold
lEyeScoreAbove = posePtr[14, 2] > threshold
rEyeScoreAbove = posePtr[15, 2] > threshold
counter = 0.2
# Leftmost point
if lEarScoreAbove:
l_true = posePtr[16, 0:2]
# Rightmost point
if rEarScoreAbove:
r_true = posePtr[17, 0:2]
counter = 0.0
if headNoseScoreAbove:
if (lEyeScoreAbove == lEarScoreAbove
and rEyeScoreAbove == rEarScoreAbove
and lEyeScoreAbove != rEyeScoreAbove):
if lEyeScoreAbove:
point_top_left += (posePtr[14, 0:2] + posePtr[16, 0:2] +
posePtr[0, 0:2]) / 3.0
face_size[1] += 0.85 * (
dist(posePtr[14, 0:2], posePtr[16, 0:2]) +
dist(posePtr[0, 0:2], posePtr[16, 0:2]) +
dist(posePtr[14, 0:2], posePtr[0, 0:2]))
face_size[0] += 0.85 * (
dist(posePtr[14, 0:2], posePtr[16, 0:2]) +
dist(posePtr[0, 0:2], posePtr[14, 0:2]))
else:
point_top_left += (posePtr[15, 0:2] + posePtr[17, 0:2] +
posePtr[0, 0:2]) / 3.0
face_size[1] += 0.85 * (
dist(posePtr[15, 0:2], posePtr[17, 0:2]) +
dist(posePtr[0, 0:2], posePtr[17, 0:2]) +
dist(posePtr[15, 0:2], posePtr[0, 0:2]))
face_size[0] += 0.85 * (
dist(posePtr[15, 0:2], posePtr[17, 0:2]) +
dist(posePtr[0, 0:2], posePtr[15, 0:2]))
counter += 1.0
elif neckScoreAbove:
point_top_left += posePtr[0, 0:2]
face_size[1] += 1.0 * dist(posePtr[1, 0:2], posePtr[0, 0:2])
face_size[0] += 0.0 * dist(posePtr[1, 0:2], posePtr[0, 0:2])
counter += 1.0
if lEyeScoreAbove and rEyeScoreAbove:
point_top_left += (posePtr[14, 0:2] + posePtr[15, 0:2]) / 2.0
face_size[1] += 4.0 * dist(posePtr[14, 0:2], posePtr[15, 0:2])
face_size[0] += 3.0 * dist(posePtr[14, 0:2], posePtr[15, 0:2])
counter += 1.0
if lEarScoreAbove and rEarScoreAbove:
point_top_left += (posePtr[16, 0:2] + posePtr[17, 0:2]) / 2.0
face_size[1] += 1.2 * dist(posePtr[16, 0:2], posePtr[17, 0:2])
face_size[0] += dist(posePtr[16, 0:2], posePtr[17, 0:2])
counter += 1.0
if counter > 0:
point_top_left /= counter
face_size[0] /= counter
face_size[1] /= counter
if l_true is not None:
ll = l_true[0]
face_size[0] = face_size[0] - (ll -
(point_top_left[0] - face_size[0] / 2))
else:
ll = point_top_left[0] - face_size[0] / 2
#
if r_true is not None:
rr = r_true[0]
face_size[0] = rr - ll
return region.RectangleRegion(
ll,
point_top_left[1] - face_size[1] / 2,
face_size[0],
face_size[1],
score,
)
