def cropBox(img, ul, br, resH, resW):
ul = ul.int()
br = br.int()
lenH = max(br[1] - ul[1], (br[0] - ul[0]) * resH / resW)
lenW = lenH * resW / resH
if img.dim() == 2:
img = img[np.newaxis, :]
newDim = torch.IntTensor((img.size(0), int(lenH), int(lenW)))
newImg = img[:, ul[1]:, ul[0]:].clone()
# Crop and Padding
size = torch.IntTensor((int(br[1] - ul[1]), int(br[0] - ul[0])))
newImg = SpecialCrop(size, 1)(newImg)
newImg = Pad(newDim)(newImg)
# Resize to output
v_Img = torch.autograd.Variable(newImg)
v_Img = torch.unsqueeze(v_Img, 0)
# newImg = F.upsample_bilinear(v_Img, size=(int(resH), int(resW))).data[0]
if torch.__version__ == '0.4.0a0+32f3bf7' or torch.__version__ == '0.4.0':
newImg = F.upsample(v_Img,
size=(int(resH), int(resW)),
mode='bilinear',
align_corners=True).data[0]
else:
newImg = F.interpolate(v_Img,
size=(int(resH), int(resW)),
mode='bilinear',
align_corners=True).data[0]
return newImg
def cropBox(img, up_left, bottom_right, target_height, target_width):
up_left = up_left.int()
bottom_right = bottom_right.int()
curr_height = max(bottom_right[1] - up_left[1],
(bottom_right[0] - up_left[0]) * target_height /
target_width)
curr_width = curr_height * target_width / target_height
if img.dim() == 2:
img = img[np.newaxis, :]
new_dim = torch.IntTensor((img.size(0), int(curr_height), int(curr_width)))
new_img = img[:, up_left[1]:, up_left[0]:]
# crop and padding
size = torch.IntTensor(
(bottom_right[1] - up_left[1], bottom_right[0] - up_left[0]))
new_img = SpecialCrop(size, 1)(new_img)
new_img = Pad(new_dim)(new_img)
# resize to output
v_img = torch.unsqueeze(new_img, 0)
# newImg = F.upsample_bilinear(v_Img, size=(int(resH), int(resW))).data[0]
new_img = F.upsample(v_img,
size=(int(target_height), int(target_width)),
mode='bilinear',
align_corners=True).data[0]
return new_img
def data_resize(img, cfg, points=None):
img = np.transpose(img, (2, 0, 1))
img_H = img.shape[1]
img_W = img.shape[2]
input_H = int(max(img_H, img_H * cfg.input_H / cfg.input_W))
input_W = int(input_H * cfg.input_W / cfg.input_H)
newDim = torch.IntTensor((img.shape[0], input_H, input_W))
newimg = torch.FloatTensor(img, )
newimg = Pad(newDim)(newimg)
v_img = torch.unsqueeze(newimg, 0)
newimg = F.upsample(v_img,
size=(cfg.input_H, cfg.input_W),
mode='bilinear',
align_corners=True).data[0]
if input_H > img_H:
top_pad_len = (input_H - img_H) / 2
if points is not None:
points[:, 1] += top_pad_len
else:
left_pad_len = (input_W - img_W) / 2
if points is not None:
points[:, 0] += left_pad_len
h_scale = cfg.input_H / input_H
w_scale = cfg.input_W / input_W
if points is not None:
points[:, 0] *= w_scale
points[:, 1] *= h_scale
return newimg, points
return newimg
def Pad_setup():
tforms = {}
tforms['pad_4040'] = Pad((40, 40))
tforms['pad_3040'] = Pad((30, 40))
tforms['pad_4030'] = Pad((40, 30))
tforms['pad_3939'] = Pad((39, 39))
tforms['pad_3941'] = Pad((39, 41))
tforms['pad_4139'] = Pad((41, 39))
tforms['pad_4138'] = Pad((41, 38))
tforms['pad_3841'] = Pad((38, 41))
return tforms
def cropBox(img, ul, br, resH, resW):
ul = ul.int() # xmin, ymin
br = br.int() # xmax, ymax
lenH = max(br[1] - ul[1], (br[0] - ul[0]) * resH / resW)
lenW = lenH * resW / resH
if img.dim() == 2:
img = img[np.newaxis, :]
newDim = torch.IntTensor((img.size(0), int(lenH), int(lenW)))
newImg = img[:, ul[1]:, ul[0]:]
# Crop and Padding
size = torch.IntTensor((br[1] - ul[1], br[0] - ul[0]))
newImg = SpecialCrop(size, 1)(newImg)
newImg = Pad(newDim)(newImg)
# Resize to output
# batch = 0, channel = img.size(0), h = int(lenH), w = int(lenW)
v_Img = torch.unsqueeze(newImg, 0)
# newImg = F.upsample_bilinear(v_Img, size=(int(resH), int(resW))).data[0]
newImg = F.upsample(v_Img, size=(int(resH), int(resW)),
mode='bilinear', align_corners=True).data[0]
return newImg