def preprocess_frame(self, frame, frame_number, labels, projection):
frame_file = self.__frame_filename(frame_number)
if frame_number < self.landmarks_df.shape[0]:
left_eye = (self.landmarks_df.iloc[[frame_number
]][0][frame_number],
self.landmarks_df.iloc[[frame_number
]][1][frame_number])
right_eye = (self.landmarks_df.iloc[[frame_number
]][2][frame_number],
self.landmarks_df.iloc[[frame_number
]][3][frame_number])
else:
return False, 'Frame number > number of landmark file rows', None
if all(v == 0 for v in (left_eye + right_eye)):
return False, 'Tracker fail: no landmarks available', None
if labels[-1]:
return False, 'Tracker fail: human labeled failure', None
cropped = np.array(
crop_face(Image.fromarray(frame),
eye_left=left_eye,
eye_right=right_eye,
offset_pct=(0.25, 0.25),
dest_sz=(64, 64)))
# gray_image = cv2.cvtColor(cropped, cv2.COLOR_BGR2GRAY)
cv2.imwrite(frame_file, cropped) # save frame as PNG file
return True, 'Success', cv2.imread(frame_file)
def preprocess_frame(self, frame, frame_number, labels):
frame_file = self.__frame_filename(frame_number)
if frame_number < self.landmarks_df.shape[0]:
left_eye = (
self.landmarks_df.iloc[[frame_number]][0][frame_number],
self.landmarks_df.iloc[[frame_number]][1][frame_number])
right_eye = (
self.landmarks_df.iloc[[frame_number]][2][frame_number],
self.landmarks_df.iloc[[frame_number]][3][frame_number])
else:
return False, 'Frame number > number of landmark file rows', None
if all(v == 0 for v in (left_eye + right_eye)):
return False, 'Tracker fail: no landmarks available', None
if labels[-1]:
return False, 'Tracker fail: human labeled failure', None
cropped = np.array(crop_face(Image.fromarray(frame), eye_left=left_eye, eye_right=right_eye,
offset_pct=(0.25, 0.25), dest_sz=(64, 64)))
to_save = cv2.cvtColor(cropped, cv2.COLOR_BGR2GRAY) if self.grayscale else cropped
cv2.imwrite(frame_file, to_save) # save frame as PNG file
return True, 'Success', cv2.imread(frame_file)
def representative_dataset_gen():
wflw_dir = "/home/robin/data/facial-marks/wflw/WFLW_images"
ds_wflw = fmd.wflw.WFLW(False, "wflw_test")
ds_wflw.populate_dataset(wflw_dir)
for _ in range(100):
sample = ds_wflw.pick_one()
# Get image and marks.
image = sample.read_image()
marks = sample.marks
# Crop the face out of the image.
image_cropped, _, _ = crop_face(image, marks, scale=1.2)
# Get the prediction from the model.
image_cropped = cv2.resize(image_cropped, (256, 256))
img_rgb = cv2.cvtColor(image_cropped, cv2.COLOR_BGR2RGB)
img_input = normalize(np.array(img_rgb, dtype=np.float32))
yield [np.expand_dims(img_input, axis=0)]
def evaluate(dataset: fmd.mark_dataset.dataset, model, n_points):
"""Evaluate the model on the dataset. The evaluation method should be the
same with the official code.
Args:
dataset: a FMD dataset
model: any model having `predict` method.
"""
# For NME
nme_count = 0
nme_sum = 0
count_failure_008 = 0
count_failure_010 = 0
# Loop though the dataset samples.
for sample in tqdm(dataset):
# Get image and marks.
image = sample.read_image()
marks = sample.marks[:n_points]
# Crop the face out of the image.
image_cropped, border, bbox = crop_face(image, marks, scale=1)
image_size = image_cropped.shape[:2]
# Get the prediction from the model.
image_cropped = cv2.resize(image_cropped, (128, 128))
img_rgb = cv2.cvtColor(image_cropped, cv2.COLOR_BGR2RGB)
# img_input = normalize(np.array(img_rgb, dtype=np.float32))
# Do prediction.
heatmaps = model.predict(tf.expand_dims(img_rgb, 0))
marks_prediction = np.reshape(heatmaps, (-1, 2))
marks_prediction *= (bbox[2] - bbox[0])
marks_prediction[:, 0] += bbox[0]
marks_prediction[:, 1] += bbox[1]
# draw_marks(image, marks_prediction)
# print(bbox)
# cv2.imwrite("sample.jpg", image)
# quit()
# Parse the heatmaps to get mark locations.
# marks_prediction, _ = parse_heatmaps(heatmaps, image_size)
# # Transform the marks back to the original image dimensions.
# x0 = bbox[0] - border
# y0 = bbox[1] - border
# marks_prediction[:, 0] += x0
# marks_prediction[:, 1] += y0
# Compute NME.
nme_temp = compute_nme(marks_prediction, marks[:, :2], n_points)
if nme_temp > 0.08:
count_failure_008 += 1
if nme_temp > 0.10:
count_failure_010 += 1
nme_sum += nme_temp
nme_count = nme_count + 1
# # Visualize the result.
# for mark in marks_prediction:
# cv2.circle(image, tuple(mark.astype(int)), 2, (0, 255, 0), -1)
# cv2.imshow("cropped", image_cropped)
# cv2.imshow("image", image)
# if cv2.waitKey(1) == 27:
# break
# NME
nme = nme_sum / nme_count
failure_008_rate = count_failure_008 / nme_count
failure_010_rate = count_failure_010 / nme_count
msg = "NME:{:.4f}, [008]:{:.4f}, [010]:{:.4f}".format(
100 * nme, failure_008_rate, failure_010_rate)
return msg