def load_process_img(self, img_idx):
imname = str(self.filenames[img_idx])
slice_idx = self.slice_idx[img_idx]
spacing = self.spacing[img_idx]
slice_intv = self.slice_intv[img_idx]
box = self.boxes[img_idx]
im, im_scale, crop = load_prep_img(imname, slice_idx, spacing, slice_intv,
do_clip=False, num_slice=config.NUM_SLICES)
box *= im_scale
patch, new_box, patch_scale = get_patch(im, box)
return patch, new_box
def get_input(nifti_data, box0, imname, infos):
slice_idx = box0[0]
box = box0[1:].copy()
vol = nifti_data.get_data()
aff = nifti_data.get_affine()[:3, :3]
spacing = np.abs(aff[:2, :2]).max()
slice_intv = np.abs(aff[2, 2])
# Ad-hoc code for normalizing the orientation of the volume.
# The aim is to make vol[:,:,i] an supine right-left slice
# It works for the authors' data, but maybe not suitable for some kinds of nifti files
if np.abs(aff[0, 0]) > np.abs(aff[0, 1]):
vol = np.transpose(vol, (1, 0, 2))
aff = aff[[1, 0, 2], :]
if np.max(aff[0, :2]) > 0:
vol = vol[::-1, :, :]
if np.max(aff[1, :2]) > 0:
vol = vol[:, ::-1, :]
im, im_scale, c = load_prep_img(vol,
int(slice_idx),
spacing=spacing,
slice_intv=slice_intv,
do_clip=False,
num_slice=config.NUM_SLICES)
box *= im_scale
patch, center_box, patch_scale = get_patch(im, box)
im_show = windowing(
windowing_rev(patch[:, :, 1] + config.PIXEL_MEANS, config.WINDOWING),
[-175, 275]).astype(np.uint8)
im_show = cv2.cvtColor(im_show, cv2.COLOR_GRAY2BGR)
cv2.rectangle(im_show, (int(center_box[0]), int(center_box[1])),
(int(center_box[2]), int(center_box[3])),
color=(0, 255, 0),
thickness=1)
coord_str = '_'.join([str(x) for x in box0.astype(int).tolist()])
fn = os.path.join(default.res_path, '%s_%s.png' % (imname, coord_str))
cv2.imwrite(fn, im_show)
infos.append('Image patch saved to %s.' % fn)
patch = patch.astype(float) / 255
patch = torch.from_numpy(patch.transpose((2, 0, 1)))
out_box = torch.tensor([[0, 0, patch.shape[2],
patch.shape[1]]]).to(dtype=torch.float).cuda()
center_box = torch.tensor(center_box[None, :]).to(dtype=torch.float).cuda()
patch = patch[None, :, :, :].to(dtype=torch.float).cuda()
return patch, out_box, center_box
def get_image(roidb):
"""
preprocess image and return processed roidb
:param roidb: a list of roidb
:return: list of img as in mxnet format
roidb add new item['im_info']
0 --- x (width, second dim of im)
|
y (height, first dim of im)
"""
num_images = len(roidb)
processed_ims = []
processed_roidb = []
# t = Timer()
for i in range(num_images):
r = roidb[i]
# t.tic()
im, im_scale, crop = load_prep_img(r['image'],
r['slice_no'],
r['spacing'],
r['slice_intv'],
config.IMG_DO_CLIP,
num_slice=config.NUM_SLICES *
config.NUM_IMAGES_3DCE)
sys.stdout.flush()
# print i, t.toc()
# print r['image']
im -= config.PIXEL_MEANS
new_r = r.copy()
im_tensor, _ = im_list_to_blob([im])
processed_ims.append(im_tensor)
im_info = [im_tensor.shape[2], im_tensor.shape[3], im_scale]
new_r['boxes'] = r['boxes'].copy()
# first clip then resize
if config.IMG_DO_CLIP:
new_r['boxes'] -= [crop[2], crop[0], crop[2], crop[0]]
new_r['crop'] = crop
new_r['boxes'] *= im_scale
new_r['im_info'] = im_info
processed_roidb.append(new_r)
if DEBUG:
from ..core.tester import vis_all_boxes
vis_all_boxes(im, new_r['boxes'])
return processed_ims, processed_roidb
def get_input(nifti_data, box0, imname, infos):
slice_idx = box0[0]
box = box0[1:].copy()
vol = nifti_data.get_data()
aff = nifti_data.get_affine()[:3, :3]
spacing = np.abs(aff[:2, :2]).max()
slice_intv = np.abs(aff[2, 2])
# Ad-hoc code for normalizing the orientation of the volume.
# The aim is to make vol[:,:,i] an supine right-left slice
# It works for the authors' data, but maybe not suitable for some kinds of nifti files
if np.abs(aff[0, 0]) > np.abs(aff[0, 1]):
vol = np.transpose(vol, (1,0,2))
aff = aff[[1,0,2], :]
if np.max(aff[0, :2]) > 0:
vol = vol[::-1, :, :]
if np.max(aff[1, :2]) > 0:
vol = vol[:, ::-1, :]
im, im_scale, c = load_prep_img(vol, int(slice_idx), spacing=spacing,
slice_intv=slice_intv, do_clip=False, num_slice=config.NUM_SLICES)
box *= im_scale
patch, center_box, patch_scale = get_patch(im, box)
im_show = windowing(windowing_rev(patch[:,:,1] + config.PIXEL_MEANS, config.WINDOWING), [-175, 275]).astype(np.uint8)
im_show = cv2.cvtColor(im_show, cv2.COLOR_GRAY2BGR)
cv2.rectangle(im_show, (int(center_box[0]), int(center_box[1])),
(int(center_box[2]), int(center_box[3])), color=(0, 255, 0), thickness=1)
coord_str = '_'.join([str(x) for x in box0.astype(int).tolist()])
fn = os.path.join(default.res_path, '%s_%s.png' % (imname, coord_str))
cv2.imwrite(fn, im_show)
infos.append('Image patch saved to %s.' % fn)
patch = patch.astype(float) / 255
patch = torch.from_numpy(patch.transpose((2, 0, 1)))
out_box = torch.tensor([[0, 0, patch.shape[2], patch.shape[1]]]).to(dtype=torch.float).cuda()
center_box = torch.tensor(center_box[None,:]).to(dtype=torch.float).cuda()
patch = patch[None, :,:,:].to(dtype=torch.float).cuda()
return patch, out_box, center_box