def preprocess(image, joint, camera, param):
logger.debug("rotate")
# random rotate
image, joint = rotate_hand(image, joint, camera, angle=param["angle"])
if param["add_noise"]:
noise = np.random.normal(0, 1, joint.shape)
joint = joint + noise
# crop
logger.debug("crop")
image, camera, domain = crop_handwith2d(
image,
joint,
camera,
param["oscillation"],
)
logger.debug("scale")
image, camera = scale(image, camera, param["crop2d"])
logger.debug("flip")
# random flip
image, joint = flip_hand(
image, joint, camera,
y_flip=param["y_flip"],
x_flip=param["x_flip"],
)
logger.debug("flip")
# scale image so that target image contains a specific domain
image, camera = resize_contain(image, camera, size=param["crop2d"], fill=0)
if param["do_color_augmentation"]:
logger.debug("distort image")
image = random_distort(image)
return image, joint, camera, domain
def __getitem__(self, _idx):
_camera, _img, _skel_2d, _skel3d, _side = super(DataPose,
self).__getitem__(_idx)
_bbox = self._getbbox(_skel_2d)
_target = self._gettarget()
_trans, _crop = self._gen_trans(_bbox, _target, _img)
if random.random() <= 0.5: # flip
_skel_2d[:, 0] = _img.shape[1] - _skel_2d[:, 0]
_skel3d[:, 0] = -_skel3d[:, 0]
_side = 1 - _side
_bbox = self._getbbox(_skel_2d)
_trans, _ = self._gen_trans(_bbox, _target, None)
_crop = _crop[:, ::-1, :]
_uv2d = self._get_point2d(_skel_2d, _trans)
_crop_img = _crop
_crop_img = _crop_img.transpose((2, 0, 1))
from chainercv.links.model.ssd import random_distort
_crop_img = random_distort(_crop_img)
if False:
_crop_img_cop = _crop_img.transpose((1, 2, 0))
self._vis(_crop_img_cop, _uv2d, _skel3d)
return np.array(_crop_img, dtype=np.float32), \
np.array(_uv2d / self.config.input_size, dtype=np.float32), \
np.uint8(_side)
def __call__(self, in_data):
img, bbox, label = in_data
# 1. Color augumentation
img = random_distort(img)
# 2. Random expansion
if np.random.randint(2):
img, param = transforms.random_expand(
img, fill=self.mean, return_param=True)
bbox = transforms.translate_bbox(
bbox, y_offset=param["y_offset"], x_offset=param["x_offset"])
# 3. Random cropping
img, param = random_crop_with_bbox_constraints(
img, bbox, return_param=True)
bbox, param = transforms.crop_bbox(
bbox, y_slice=param["y_slice"], x_slice=param["x_slice"],
allow_outside_center=False, return_param=True)
label = label[param["index"]]
# 4. Resizing with random interpolation
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Transformation for SSD network input
img -= self.mean
mb_loc, mb_lab = self.coder.encode(bbox, label)
return img, mb_loc, mb_lab
def __call__(self, in_data):
img, bbox, label = in_data
# 1. Color augumentation
img = random_distort(img)
# 2. Random expansion
if np.random.randint(2):
img, param = transforms.random_expand(
img, fill=self.mean, return_param=True)
bbox = transforms.translate_bbox(
bbox, y_offset=param["y_offset"], x_offset=param["x_offset"])
# 3. Random cropping
img, param = random_crop_with_bbox_constraints(
img, bbox, return_param=True)
bbox, param = transforms.crop_bbox(
bbox, y_slice=param["y_slice"], x_slice=param["x_slice"],
allow_outside_center=False, return_param=True)
label = label[param["index"]]
# 4. Resizing with random interpolation
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Transformation for SSD network input
img -= self.mean
mb_loc, mb_lab = self.coder.encode(bbox, label)
return img, mb_loc, mb_lab
def test_random_distort(self):
if not optional_modules:
return
img = np.random.randint(0, 256, size=(3, 48, 32)).astype(np.float32)
out = random_distort(img)
self.assertEqual(out.shape, img.shape)
self.assertEqual(out.dtype, img.dtype)
def __call__(self, in_data):
# There are five data augmentation steps
# 1. Color augmentation
# 2. Random expansion
# 3. Random cropping
# 4. Resizing with random interpolation
# 5. Random horizontal flipping
# 6. Random vertical flipping
img, bbox, label = in_data
# 1. Color augmentation
img = random_distort(img)
# 2. Random expansion
if np.random.randint(2):
img, param = transforms.random_expand(img,
fill=self.mean,
return_param=True)
bbox = transforms.translate_bbox(bbox,
y_offset=param['y_offset'],
x_offset=param['x_offset'])
# 3. Random cropping
img, param = random_crop_with_bbox_constraints(img,
bbox,
return_param=True)
bbox, param = transforms.crop_bbox(bbox,
y_slice=param['y_slice'],
x_slice=param['x_slice'],
allow_outside_center=False,
return_param=True)
label = label[param['index']]
# 4. Resizing with random interpolatation
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Random horizontal flipping
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox, (self.size, self.size),
x_flip=params['x_flip'])
# 6. Random vertical flipping
img, params = transforms.random_flip(img,
y_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox, (self.size, self.size),
y_flip=params['y_flip'])
# Preparation for SSD network
img -= self.mean
mb_loc, mb_label = self.coder.encode(bbox, label)
return img, mb_loc, mb_label
def __call__(self, in_data):
# 5段階のステップでデータの水増しを行う
# 1. 色の拡張
# 2. ランダムな拡大
# 3. ランダムなトリミング
# 4. ランダムな補完の再補正
# 5. ランダムな水平反転
img, bbox, label = in_data
# 1. 色の拡張
# 明るさ，コントラスト，彩度，色相を組み合わせ，データ拡張をする
img = random_distort(img)
# 2. ランダムな拡大
if np.random.randint(2):
# キャンバスの様々な座標に入力画像を置いて，様々な比率の画像を生成し，bounding boxを更新
img, param = transforms.random_expand(img,
fill=self.mean,
return_param=True)
bbox = transforms.translate_bbox(bbox,
y_offset=param['y_offset'],
x_offset=param['x_offset'])
# 3. ランダムなトリミング
img, param = random_crop_with_bbox_constraints(img,
bbox,
return_param=True)
# トリミングされた画像内にbounding boxが入るように調整
bbox, param = transforms.crop_bbox(bbox,
y_slice=param['y_slice'],
x_slice=param['x_slice'],
allow_outside_center=False,
return_param=True)
label = label[param['index']]
# 4. ランダムな補完の再補正
## 画像とbounding boxのリサイズ
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. ランダムな水平反転
## 画像とbounding boxをランダムに水平方向に反転
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox, (self.size, self.size),
x_flip=params['x_flip'])
# SSDのネットワークに入力するための準備の処理
img -= self.mean
## SSDに入力するためのloc(デフォルトbounding boxのオフセットとスケール)と
## mb_label(クラスを表す配列)を出力
mb_loc, mb_label = self.coder.encode(bbox, label)
return img, mb_loc, mb_label
def __call__(self, in_data):
# There are five data augmentation steps
# 1. Color augmentation
# 2. Random expansion
# 3. Random cropping
# 4. Resizing with random interpolation
# 5. Random horizontal flipping
img, bbox, label = in_data
bbox = np.array(bbox).astype(np.float32)
if len(bbox) == 0:
warnings.warn("No bounding box detected", RuntimeWarning)
img = resize_with_random_interpolation(img, (self.size, self.size))
mb_loc, mb_label = self.coder.encode(bbox, label)
return img, mb_loc, mb_label
# 1. Color augmentation
img = random_distort(img)
# 2. Random expansion
if np.random.randint(2):
img, param = transforms.random_expand(
img, fill=self.mean, return_param=True)
bbox = transforms.translate_bbox(
bbox, y_offset=param['y_offset'], x_offset=param['x_offset'])
# 3. Random cropping
img, param = random_crop_with_bbox_constraints(
img, bbox, return_param=True)
bbox, param = transforms.crop_bbox(
bbox, y_slice=param['y_slice'], x_slice=param['x_slice'],
allow_outside_center=False, return_param=True)
label = label[param['index']]
# 4. Resizing with random interpolatation
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Random horizontal flipping
img, params = transforms.random_flip(
img, x_random=True, return_param=True)
bbox = transforms.flip_bbox(
bbox, (self.size, self.size), x_flip=params['x_flip'])
mb_loc, mb_label = self.coder.encode(bbox, label)
return img, mb_loc, mb_label
def __call__(self, in_data):
# There are five data augmentation steps
# 1. Color augmentation
# 2. Random expansion
# 3. Random cropping
# 4. Resizing with random interpolation
# 5. Random horizontal flipping
img, bbox, label = in_data
# 1. Color augmentation
img = random_distort(img)
# 2. Random expansion
if np.random.randint(2):
img, param = transforms.random_expand(
img, fill=self.mean, return_param=True)
bbox = transforms.translate_bbox(
bbox, y_offset=param['y_offset'], x_offset=param['x_offset'])
# 3. Random cropping
img, param = random_crop_with_bbox_constraints(
img, bbox, return_param=True)
bbox, param = transforms.crop_bbox(
bbox, y_slice=param['y_slice'], x_slice=param['x_slice'],
allow_outside_center=False, return_param=True)
label = label[param['index']]
# 4. Resizing with random interpolatation
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Random horizontal flipping
img, params = transforms.random_flip(
img, x_random=True, return_param=True)
bbox = transforms.flip_bbox(
bbox, (self.size, self.size), x_flip=params['x_flip'])
# Preparation for SSD network
img -= self.mean
mb_loc, mb_label = self.coder.encode(bbox, label)
return img, mb_loc, mb_label
def __call__(self, in_data):
img, bbox, label = in_data
img = random_distort(img)
if np.random.randint(2):
img, param = transforms.random_expand(img,
fill=self.mean,
return_param=True)
bbox = transforms.translate_bbox(bbox,
y_offset=param['y_offset'],
x_offset=param['x_offset'])
img, param = random_crop_with_bbox_constraints(img,
bbox,
return_param=True)
bbox, param = transforms.crop_bbox(bbox,
y_slice=param['y_slice'],
x_slice=param['x_slice'],
allow_outside_center=False,
return_param=True)
label = label[param['index']]
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox, (self.size, self.size),
x_flip=params['x_flip'])
img -= self.mean
mb_loc, mb_label = self.coder.encode(bbox, label)
return img, mb_loc, mb_label
def __call__(self, in_data):
"""in_data includes three datas.
Args:
img(array): Shape is (3, H, W). range is [0, 255].
bbox(array): Shape is (N, 4). (y_min, x_min, y_max, x_max).
range is [0, max size of boxes].
label(array): Classes of bounding boxes.
Returns:
img(array): Shape is (3, out_H, out_W). range is [0, 1].
interpolation value equals to self.value.
"""
# There are five data augmentation steps
# 1. Color augmentation
# 2. Random expansion
# 3. Random cropping
# 4. Resizing with random interpolation
# 5. Random horizontal flipping
if self.count % 10 == 0 and self.count % self.batchsize == 0 and self.count != 0:
self.i += 1
i = self.i % len(self.dim)
self.output_shape = (self.dim[i], self.dim[i])
self.count += 1
img, bbox, label = in_data
# 1. Color augmentation
img = random_distort(img,
brightness_delta=32,
contrast_low=0.5,
contrast_high=1.5,
saturation_low=0.5,
saturation_high=1.5,
hue_delta=25)
# Normalize. range is [0, 1]
img /= 255.0
_, H, W = img.shape
scale = np.random.uniform(0.25, 2)
random_expand = np.random.uniform(0.8, 1.2, 2)
net_h, net_w = self.output_shape
out_h = net_h * scale # random_expand[0]
out_w = net_w * scale # random_expand[1]
if H > W:
out_w = out_h * (float(W) / H) * np.random.uniform(0.8, 1.2)
elif H < W:
out_h = out_w * (float(H) / W) * np.random.uniform(0.8, 1.2)
out_h = int(out_h)
out_w = int(out_w)
img = resize_with_random_interpolation(img, (out_h, out_w))
bbox = transforms.resize_bbox(bbox, (H, W), (out_h, out_w))
if out_h < net_h and out_w < net_w:
img, param = expand(img,
out_h=net_h,
out_w=net_w,
fill=self.value,
return_param=True)
bbox = transforms.translate_bbox(bbox,
y_offset=param['y_offset'],
x_offset=param['x_offset'])
else:
out_h = net_h if net_h > out_h else int(out_h * 1.05)
out_w = net_w if net_w > out_w else int(out_w * 1.05)
img, param = expand(img,
out_h=out_h,
out_w=out_w,
fill=self.value,
return_param=True)
bbox = transforms.translate_bbox(bbox,
y_offset=param['y_offset'],
x_offset=param['x_offset'])
img, param = crop_with_bbox_constraints(img,
bbox,
return_param=True,
crop_height=net_h,
crop_width=net_w)
bbox, param = transforms.crop_bbox(bbox,
y_slice=param['y_slice'],
x_slice=param['x_slice'],
allow_outside_center=False,
return_param=True)
label = label[param['index']]
# 5. Random horizontal flipping # OK
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox,
self.output_shape,
x_flip=params['x_flip'])
# Preparation for Yolov2 network. scale=[0, 1]
bbox[:, ::2] /= self.output_shape[0] # y
bbox[:, 1::2] /= self.output_shape[1] # x
num_bbox = len(bbox)
len_max = max(num_bbox, self.max_target)
out_bbox = np.zeros((len_max, 4), dtype='f')
out_bbox[:num_bbox] = bbox[:num_bbox]
out_label = np.zeros((len_max), dtype='i')
out_label[:num_bbox] = label
out_bbox = out_bbox[:self.max_target]
out_label = out_label[:self.max_target]
num_array = min(num_bbox, self.max_target)
gmap = create_map_anchor_gt(bbox, self.anchors, self.output_shape,
self.downscale, self.n_boxes, len_max)
gmap = gmap[:self.max_target]
img = np.clip(img, 0, 1)
return img, out_bbox, out_label, gmap, np.array([num_array], dtype='i')
def _data_augumentation(self, image):
image = random_distort(image)
image = random_flip(image, x_random=True)
return image
def __getitem__(self, index):
img_path = self.img_files[index % len(self.img_files)].rstrip()
img = np.array(Image.open(img_path))
h, w, _ = img.shape
label_path = self.label_files[index % len(self.img_files)].rstrip()
if not os.path.exists(label_path):
raise Exception(
"the label file(.txt) is not found corresponding + " +
img_path)
labels = np.loadtxt(label_path).reshape(-1, 5)
#--------------------
# data augmentation
#--------------------
lx = w * (labels[:, 1] - labels[:, 3] / 2)
ly = h * (labels[:, 2] - labels[:, 4] / 2)
bx = w * (labels[:, 1] + labels[:, 3] / 2)
by = h * (labels[:, 2] + labels[:, 4] / 2)
# convert to chainercv format: (ly, lx, by, bx)
cv_bbox = np.stack([ly, lx, by, bx], axis=1)
cv_labels = labels[:, 0].reshape(-1).astype(np.int)
cv_img = img.transpose(2, 0, 1)
# 1. Random distort
cv_img = random_distort(cv_img)
# 2. Random cropping
cv_img, param = random_crop_with_bbox_constraints(cv_img,
cv_bbox,
min_scale=0.3,
return_param=True)
cv_bbox, param = transforms.crop_bbox(cv_bbox,
y_slice=param['y_slice'],
x_slice=param['x_slice'],
allow_outside_center=False,
return_param=True)
cv_labels = cv_labels[param['index']]
# 3. Random horizontal flipping
_, _h, _w = cv_img.shape
cv_img, params = transforms.random_flip(cv_img,
x_random=True,
return_param=True)
cv_bbox = transforms.flip_bbox(cv_bbox, (_h, _w),
x_flip=params['x_flip'])
# update params
img = cv_img.transpose(1, 2, 0)
h, w, _ = img.shape
# convert to default format: (cx, cy, w, h)
labels = np.zeros((cv_labels.size, 5))
labels[:, 0] = cv_labels
labels[:, 1] = (cv_bbox[:, 3] + cv_bbox[:, 1]) / 2.0 / w # cx
labels[:, 2] = (cv_bbox[:, 2] + cv_bbox[:, 0]) / 2.0 / h # cy
labels[:, 3] = (cv_bbox[:, 3] - cv_bbox[:, 1]) / w # w
labels[:, 4] = (cv_bbox[:, 2] - cv_bbox[:, 0]) / h # x
#---------
# image
#---------
dim_diff = np.abs(h - w)
# Upper (left) and lower (right) padding
pad1, pad2 = dim_diff // 2, dim_diff - dim_diff // 2
# Determine padding
pad = ((pad1, pad2), (0, 0),
(0, 0)) if h <= w else ((0, 0), (pad1, pad2), (0, 0))
# Add padding
input_img = np.pad(img, pad, 'constant', constant_values=128) / 255.
padded_h, padded_w, _ = input_img.shape
# Resize and normalize
input_img = resize(input_img, (*self.img_shape, 3), mode='reflect')
# Channels-first
input_img = np.transpose(input_img, (2, 0, 1))
# As pytorch tensor
input_img = torch.from_numpy(input_img).float()
#---------
# Label
#---------
# Extract coordinates for unpadded + unscaled image
x1 = w * (labels[:, 1] - labels[:, 3] / 2)
y1 = h * (labels[:, 2] - labels[:, 4] / 2)
x2 = w * (labels[:, 1] + labels[:, 3] / 2)
y2 = h * (labels[:, 2] + labels[:, 4] / 2)
# Adjust for added padding
x1 += pad[1][0]
y1 += pad[0][0]
x2 += pad[1][0]
y2 += pad[0][0]
# Calculate ratios from coordinates
labels[:, 1] = ((x1 + x2) / 2) / padded_w
labels[:, 2] = ((y1 + y2) / 2) / padded_h
labels[:, 3] *= w / padded_w
labels[:, 4] *= h / padded_h
# Fill matrix
filled_labels = np.zeros((self.max_objects, 5))
if labels is not None:
filled_labels[range(
len(labels))[:self.max_objects]] = labels[:self.max_objects]
filled_labels = torch.from_numpy(filled_labels)
return img_path, input_img, filled_labels
def __call__(self, in_data):
# There are five data augmentation steps
# 1. Color augmentation
# 2. Random expansion
# 3. Random cropping
# 4. Resizing with random interpolation
# 5. Random horizontal flipping
# mask = None
img, bbox, label, mask = in_data
# TODO: show information
# self._show_img(img)
# self._show_mask(mask)
# 1. Color augmentation
img = random_distort(img)
# self._show_img(img)
# 2. Random expansion
if np.random.randint(2):
img, param = transforms.random_expand(
img, fill=self.mean, return_param=True)
bbox = transforms.translate_bbox(
bbox, y_offset=param['y_offset'], x_offset=param['x_offset'])
if mask is not None:
_, new_height, new_width = img.shape
param['new_height'] = new_height
param['new_width'] = new_width
mask = self._random_expand_mask(mask, param)
# self._show_img(img)
# self._show_mask(mask)
# 3. Random cropping
img, param = random_crop_with_bbox_constraints(
img, bbox, return_param=True)
# self._show_img(img)
mask = self._fixed_crop_mask(mask, param['y_slice'], param['x_slice'])
# self._show_mask(mask)
bbox, param = transforms.crop_bbox(
bbox, y_slice=param['y_slice'], x_slice=param['x_slice'],
allow_outside_center=False, return_param=True)
label = label[param['index']]
# 4. Resizing with random interpolatation
_, H, W = img.shape
img = resize_with_random_interpolation(img, (self.size, self.size))
# self._show_img(img)
if mask is not None:
if mask.size == 0:
raise RuntimeError
mask = self._resize_with_nearest(mask, (self.size, self.size))
# self._show_mask(mask)
bbox = transforms.resize_bbox(bbox, (H, W), (self.size, self.size))
# 5. Random horizontal flipping
img, params = transforms.random_flip(
img, x_random=True, return_param=True)
bbox = transforms.flip_bbox(
bbox, (self.size, self.size), x_flip=params['x_flip'])
if mask is not None:
mask = self._random_flip_mask(mask, x_flip=params['x_flip'], y_flip=params['y_flip'])
# self._show_img(img)
# self._show_mask(mask)
# Preparation for SSD network
img -= self.mean
mb_loc, mb_label = self.coder.encode(bbox, label)
if mask is None:
mask = np.ones([self.size, self.size], dtype=np.int32) * -1
# print("Dtype is :"+str(mask.dtype))
data_type = str(mask.dtype)
target_type = 'int32'
if data_type != target_type:
mask = mask.astype(np.int32)
if img is None:
raise RuntimeError
return img, mb_loc, mb_label, mask
