def apply_image(self, img):
'''
Resize image with random scale and position
Args:
img (PIL image): image to be resized
Return:
img (PIL image): resized image
s (dict):
ratio (tuple), scale of width and height
'''
np_img = np.array(img)
mean = np_img.mean(axis=0).mean(axis=0)
h, w = np_img.shape[0], np_img.shape[1]
in_w, in_h = self.size
c = np.array([w / 2., h / 2.], dtype=np.float32)
s = max(h, w) * 1.0
# image is randomly dragged to an center area with width and height of half of original
w_border = get_border(int(w * 0.5), w)
h_border = get_border(int(h * 0.5), h)
s = s * np.random.uniform(self.scale[0], self.scale[1]) #np.random.choice(np.arange(0.8, 1.5, 0.1))
c[0] = np.random.randint(low=w_border, high=w - w_border)
c[1] = np.random.randint(low=h_border, high=h - h_border)
trans_input = get_affine_transform(c, s, 0, [in_w, in_h])
np_img = cv2.warpAffine(np_img, trans_input, (in_w, in_h), flags=cv2.INTER_LINEAR, borderValue=(int(mean[0]), int(mean[1]), int(mean[2])))
img = Image.fromarray(np_img)
s = {'c': c, 's': s}
return img, s
def apply_pts(self, cid, pts, s):
'''
Resize keypoints with random scale and position
Args:
cid (int): the class for keypoints
pts (numpy.ndarray, shape Nx2): keypoints to be resized
s (dict):
ratio (tuple), scale of width and height recorded in function, apply_image
Return:
pts (numpy.ndarray, shape Nx2): resized keypoints
'''
assert 'c' in s
assert 's' in s
out_h, out_w = (np.array(self.size) // self.stride).astype(int)
trans_output = get_affine_transform(s['c'], s['s'], 0, [out_w, out_h])
for i in range(pts.shape[0]):
pts[i] = affine_transform(pts[i], trans_output)
return pts
def apply_bbox(self, bbox, s):
'''
Resize bbox with random scale and position
Args:
bbox (numpy.ndarray, shape 1x4): bbox to be resized
s (dict):
ratio (tuple), scale of width and height recorded in function, apply_image
Return:
bbox (numpy.ndarray, shape 1x4): resized bbox
'''
assert 'c' in s
assert 's' in s
out_w, out_h = np.array(self.size)
trans_output = get_affine_transform(s['c'], s['s'], 0, [out_w, out_h])
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:4] = affine_transform(bbox[2:4], trans_output)
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, out_w - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, out_h - 1)
return bbox
def apply_pts(self, cid, pts, s):
'''
Resize keypoints
Args:
cid (int): the class for keypoints
pts (numpy.ndarray, shape Nx2): keypoints to be resized
s (dict):
ratio (tuple), scale of width and height recorded in function, apply_image
Return:
pts (numpy.ndarray, shape Nx2): resized keypoints
'''
assert 'c' in s
assert 's' in s
out_w, out_h = np.array(self.size)
trans_output = get_affine_transform(s['c'], s['s'], 0, [out_w, out_h])
for i in range(pts.shape[0]):
pts[i,:2] = affine_transform(pts[i,:2], trans_output)
if ((pts[i, :2] < 0).sum() + (pts[i, :2] > (out_w, out_h)).sum()) > 0:
pts[i, 2] = 0.0
return pts
def apply_image(self, img):
'''
Resize image
Args:
img (PIL image): image to be resized
Return:
img (PIL image): resized image
s (dict):
ratio (tuple), scale of width and height
'''
np_img = np.array(img)
mean = np_img.mean(axis=0).mean(axis=0)
h, w = np_img.shape[0], np_img.shape[1]
in_w, in_h = self.size
c = np.array([w / 2., h / 2.], dtype=np.float32)
s = max(h, w) * 1.0
trans_input = get_affine_transform(c, s, 0, [in_w, in_h])
np_img = cv2.warpAffine(np_img, trans_input, (in_w, in_h), flags=cv2.INTER_LINEAR, borderValue=(int(mean[0]), int(mean[1]), int(mean[2])))
img = Image.fromarray(np_img)
s = {'c': c, 's': s}
return img, s