def __getitem__(self, index):
joints_data = self.data[index].copy()
# Leer la imagen del disco
image = cv2.imread(joints_data['imgPath'], cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
if self.color_rgb:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if image is None:
raise ValueError('Fail to read %s' % image)
joints = joints_data['joints']
joints_vis = joints_data['joints_visibility']
c = joints_data['center']
s = joints_data['scale']
score = joints_data['score'] if 'score' in joints_data else 1
r = 0
# Aplicar aumento de datos
if self.is_train:
if self.half_body_prob and \
random.random() < self.half_body_prob and \
np.sum(joints_vis[:, 0]) > self.nof_joints_half_body:
c_half_body, s_half_body = self._half_body_transform(joints, joints_vis)
if c_half_body is not None and s_half_body is not None:
c, s = c_half_body, s_half_body
sf = self.scale_factor
rf = self.rotation_factor
if self.scale:
s = s * np.clip(random.random() * sf + 1, 1 - sf, 1 + sf) # Un factor de escala aleatorio en [1 - sf, 1 + sf]
if self.rotate_prob and random.random() < self.rotate_prob:
r = np.clip(random.random() * rf, -rf * 2, rf * 2) # Un factor de rotación aleatorio en [-2 * rf, 2 * rf]
else:
r = 0
if self.flip_prob and random.random() < self.flip_prob:
image = image[:, ::-1, :]
joints, joints_vis = fliplr_joints(joints, joints_vis, image.shape[1], self.flip_pairs)
c[0] = image.shape[1] - c[0] - 1
#Aplicamos transformación en las articulaciones y la imagen.
trans = get_affine_transform(c, s, self.pixel_std, r, self.image_size)
image = cv2.warpAffine(
image,
trans,
(int(self.image_size[0]), int(self.image_size[1])),
flags=cv2.INTER_LINEAR
)
for i in range(self.nof_joints):
if joints_vis[i, 0] > 0.:
joints[i, 0:2] = affine_transform(joints[i, 0:2], trans)
# Convertir imagen a tensor y normalizar
if self.transform is not None:
image = self.transform(image)
target, target_weight = self._generate_target(joints, joints_vis)
# Actualizar metadatos
joints_data['joints'] = joints
joints_data['joints_visibility'] = joints_vis
joints_data['center'] = c
joints_data['scale'] = s
joints_data['rotation'] = r
joints_data['score'] = score
return image, target.astype(np.float32), target_weight.astype(np.float32), joints_data
def __getitem__(self, index):
joints_data = self.data[index].copy()
# Read the image from disk
image = cv2.imread(joints_data['imgPath'], cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
if self.color_rgb:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if image is None:
raise ValueError('Fail to read %s' % image)
joints = joints_data['joints']
joints_vis = joints_data['joints_visibility']
c = joints_data['center']
s = joints_data['scale']
score = joints_data['score'] if 'score' in joints_data else 1
r = 0
# Apply data augmentation
if self.is_train:
if self.half_body_prob and \
random.random() < self.half_body_prob and \
np.sum(joints_vis[:, 0]) > self.nof_joints_half_body:
c_half_body, s_half_body = self._half_body_transform(joints, joints_vis)
if c_half_body is not None and s_half_body is not None:
c, s = c_half_body, s_half_body
sf = self.scale_factor
rf = self.rotation_factor
if self.scale:
s = s * np.clip(random.random() * sf + 1, 1 - sf, 1 + sf) # A random scale factor in [1 - sf, 1 + sf]
if self.rotate_prob and random.random() < self.rotate_prob:
r = np.clip(random.random() * rf, -rf * 2, rf * 2) # A random rotation factor in [-2 * rf, 2 * rf]
else:
r = 0
if self.flip_prob and random.random() < self.flip_prob:
image = image[:, ::-1, :]
joints, joints_vis = fliplr_joints(joints, joints_vis, image.shape[1], self.flip_pairs)
c[0] = image.shape[1] - c[0] - 1
# Apply affine transform on joints and image
trans = get_affine_transform(c, s, self.pixel_std, r, self.image_size)
image = cv2.warpAffine(
image,
trans,
(int(self.image_size[0]), int(self.image_size[1])),
flags=cv2.INTER_LINEAR
)
for i in range(self.nof_joints):
if joints_vis[i, 0] > 0.:
joints[i, 0:2] = affine_transform(joints[i, 0:2], trans)
# Convert image to tensor and normalize
if self.transform is not None: # I could remove this check
image = self.transform(image)
target, target_weight = self._generate_target(joints, joints_vis)
# Update metadata
joints_data['joints'] = joints
joints_data['joints_visibility'] = joints_vis
joints_data['center'] = c
joints_data['scale'] = s
joints_data['rotation'] = r
joints_data['score'] = score
return image, target.astype(np.float32), target_weight.astype(np.float32), joints_data