def detect_all_faces(self):
bbox_landmark_dict = {}
for img_path in self.image_list:
print(img_path)
try:
cur_img = Image.open(img_path).convert('RGB')
except:
continue
bbox, landmark = detect_faces(cur_img)
bbox_landmark_dict[img_path] = [bbox, landmark]
return bbox_landmark_dict
def main():
method = "haar"
if len(sys.argv) == 5 and sys.argv[4] == "mtcnn":
method = "mtcnn"
f = fetch_file_for_image(sys.argv[1], sys.argv[2], sys.argv[3])
if method == "haar":
import haar
faces = haar.detect_faces(f)
upload_face_segments(sys.argv[1], sys.argv[3], faces)
else:
import mtcnn
faces = mtcnn.detect_faces(f)
upload_face_segments(sys.argv[1], sys.argv[3], faces)
os.remove(f)
def emotion():
model = Face_Emotion_CNN()
model.load_state_dict(torch.load(
'/home/cpz/Desktop/AlphaPose/emotion/models/FER_trained_model.pt',
map_location=lambda storage, loc: storage),
strict=False)
emotion_dict = {
0: 'neutral',
1: 'happiness',
2: 'surprise',
3: 'sadness',
4: 'anger',
5: 'disguest',
6: 'fear'
}
val_transform = transforms.Compose([transforms.ToTensor()])
img = cv2.imread('../input/2.jpg')
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
det_frame = Image.fromarray(img)
bounding_boxes, landmarks = detect_faces(det_frame)
for (xmin, ymin, xmax, ymax, confidence) in bounding_boxes:
ymin, xmin, ymax, xmax = int(ymin), int(xmin), int(ymax), int(xmax)
cv2.rectangle(img, (xmin, ymin), (xmax, ymax), (255, 0, 0), 2)
resize_frame = cv2.resize(gray[ymin:ymax, xmin:xmax], (48, 48))
X = resize_frame / 256
X = Image.fromarray((resize_frame))
X = val_transform(X).unsqueeze(0)
with torch.no_grad():
model.eval()
log_ps = model.cpu()(X)
ps = torch.exp(log_ps)
top_p, top_class = ps.topk(1, dim=1)
pred = emotion_dict[int(top_class.numpy())]
cv2.putText(img, pred, (xmin, ymin), cv2.FONT_HERSHEY_SIMPLEX, 0.8,
(0, 255, 0), 1)
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.grid(False)
plt.axis('off')
plt.show()
def get_features(self, img):
"""
A function which detects faces in the image, and computes the feature vector for each detection
:param img: numpy.ndarray, dimensions: (height, weight, 3)
:return: numpy.ndarray, an array of features, dimensions: (num_img, 1024)
"""
features = []
try:
bboxes, landmarks = detect_faces(img)
if len(bboxes) != 0:
faces = []
for box_landmarks in landmarks:
if img.ndim == 3:
# For some reason, the image sometimes contains all the colors and sometimes not -->
# doesn't matter as the classifier expects 2D image anyway
img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
face_img = self._frontalize_face(img, box_landmarks)
face_img = self._process_image(face_img)
faces.append(face_img)
faces = np.vstack(faces)
features = self._get_features_for_batch(faces)
except Exception as err:
print(f'\033[93mException: {err} --> skipping the frame classification\033[0m')
return features
import cv2
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--with_draw', help='do draw?', default='True')
args = parser.parse_args()
bgr_img = cv2.imread('test.jpg', 1)
print (bgr_img.shape)
### detection
list_time = []
for idx in range(10):
rgb_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
start = cv2.getTickCount()
bounding_boxes, landmarks = detect_faces(rgb_img)
time = (cv2.getTickCount() - start) / cv2.getTickFrequency() * 1000
list_time.append(time)
print ('mtcnn average time: %.3f ms'%np.array(list_time[1:]).mean())
### draw rectangle bbox
if args.with_draw == 'True':
for b in bounding_boxes:
b = [int(round(value)) for value in b]
cv2.rectangle(bgr_img, (b[0], b[1]), (b[2], b[3]), (0,255,0), 2)
for p in landmarks:
for i in range(5):
cv2.circle(bgr_img, (p[i] , p[i + 5]), 3, (255,0,0), -1)
image_key = f'name={name_formatted}&video_date={video_date}&frame={frame}'
if image_key in file_map:
# Image already exists
image_path = path.join(conf.DATASET,
name_formatted,
file_map[image_key])
region = image_path.split(
'®ion=')[1][:-4].split('_')
bbox = [float(x) for x in region]
else:
# Load the frame
im = get_frame(video, frame)
# 8) Get bboxes and landmarks
bboxes, _ = detect_faces(im)
bbox_i = get_bbox_i_by_IoU(bboxes,
rect,
threshold=0)
if bbox_i == -1:
bbox = []
else:
bbox = bboxes[bbox_i]
bbox = np.round(bbox, decimals=2)
bbox = bbox.tolist()
detection_dict['orig'] = rect
detection_dict['mtcnn'] = bbox
mtcnn_file.write(json.dumps(detection_dict) + '\n')
except Exception as e:
def mtcnn_detector(img):
bounding_boxes, _ = detect_faces(img)
faces = [(int(round(box[0])), int(round(box[1])),
int(round(box[2] - box[0])), int(round(box[3] - box[1])))
for box in bounding_boxes]
return faces
try:
# 6) Iterate over detections which belong to the name
for detection in annotations[name]['detections']:
if N_counter != N:
N_counter += 1
continue
N_counter = 1
frame = detection['frame']
rect = detection['rect']
# Load the frame
im = get_frame(video, frame)
# 8) Get bboxes and landmarks
bboxes, landmarks = detect_faces(im)
bbox_i = get_bbox_i_by_IoU(bboxes, rect)
# 9) Skip the image if no bbox was selected
if bbox_i == -1:
with open(log_path, 'a') as f:
f.write(
'{"video_name": '
f'"{video_name}", "name": "{name}", "frame": {frame}, "rect": {rect}'
'}\n')
not_found_counter += 1
continue
face_landmarks = landmarks[bbox_i]
face_bbox = np.round(bboxes[bbox_i], decimals=2)
serialized_bbox = '_'.join(str(v) for v in face_bbox)