def preprocess_inputs(self, raw_inputs, params):
outputs = dict()
# label & inst.
transform_label = get_transform_fn(self.opt,
params,
method=Image.NEAREST,
normalize=False)
outputs['label'] = transform_label(raw_inputs['label']) * 255.0
outputs['inst'] = transform_label(raw_inputs['inst'])
if self.opt.dataloader == 'sun_rgbd' or self.opt.dataloader == 'ade20k': # NOTE(sh): dirty exception!
outputs['inst'] *= 255.0
outputs['label_path'] = raw_inputs['label_path']
outputs['inst_path'] = raw_inputs['inst_path']
# image
if self.load_image:
transform_image = get_transform_fn(self.opt, params)
outputs['image'] = transform_image(raw_inputs['image'])
outputs['image_path'] = raw_inputs['image_path']
# raw inputs
if self.load_raw:
transform_raw = get_raw_transform_fn(normalize=False)
outputs['label_raw'] = transform_raw(raw_inputs['label']) * 255.0
outputs['inst_raw'] = transform_raw(raw_inputs['inst'])
transform_image_raw = get_raw_transform_fn()
outputs['image_raw'] = transform_image_raw(raw_inputs['image'])
return outputs
def preprocess_cropping(self, raw_inputs, outputs, params):
transform_obj = get_transform_fn(self.opt,
params,
method=Image.NEAREST,
normalize=False,
is_context=False)
label_obj = transform_obj(raw_inputs['label']) * 255.0
input_bbox = np.array(params['bbox_in_context'])
bbox_cls = params['bbox_cls']
bbox_cls = bbox_cls if bbox_cls is not None else self.opt.label_nc - 1
mask_object_inst = (outputs['inst']==params['bbox_inst_id']).float() \
if not (params['bbox_inst_id'] == None) else torch.zeros(outputs['inst'].size())
### generate output bbox
img_size = outputs['label'].size(1) #shape[1]
context_ratio = np.random.uniform(low=self.config['min_ctx_ratio'],
high=self.config['max_ctx_ratio'])
output_bbox = np.array(
get_soft_bbox(input_bbox, img_size, img_size, context_ratio))
mask_in, mask_object_in, mask_context_in = get_masked_image(
outputs['label'], input_bbox, bbox_cls)
mask_out, mask_object_out, _ = get_masked_image(
outputs['label'], output_bbox)
# Build dictionary
outputs['input_bbox'] = torch.from_numpy(input_bbox)
outputs['output_bbox'] = torch.from_numpy(output_bbox)
outputs['mask_in'] = mask_in # (1x1xHxW)
outputs['mask_object_in'] = mask_object_in # (1xCxHxW)
outputs['mask_context_in'] = mask_context_in # (1xCxHxW)
outputs['mask_out'] = mask_out # (1x1xHxW)
outputs['mask_object_out'] = mask_object_out # (1xCxHxW)
outputs['label_obj'] = label_obj
outputs['mask_object_inst'] = mask_object_inst
outputs['cls'] = torch.LongTensor([bbox_cls])
return outputs
def crop_canvas(bbox_sampled, label_original, opt, img_original=None, \
patch_to_obj_ratio=1.2, min_ctx_ratio=1.2, max_ctx_ratio=1.5, resize=True, \
transform_img=False):
h, w = label_original.size()[2:4]
config = {}
config['prob_flip'] = 0.0
config['fineSize'] = opt.fineSize if resize else None
config['img_to_obj_ratio'] = opt.contextMargin
config['patch_to_obj_ratio'] = patch_to_obj_ratio
config['min_ctx_ratio'] = min_ctx_ratio
config['max_ctx_ratio'] = max_ctx_ratio
params = get_transform_params((w,h), config=config, bbox=bbox_sampled, \
random_crop=False)
transform_label = get_transform_fn(opt, params, method=Image.NEAREST, \
normalize=False, resize=resize)
transform_image = get_transform_fn(opt, params, resize=resize)
output_dict = {}
output_dict['label'] = transform_label(tensor2pil(label_original[0])) * 255.0
if transform_img:
output_dict['image'] = transform_image(tensor2pil(img_original[0], \
is_img=True))
input_bbox = np.array(params['bbox_in_context'])
crop_pos = np.array(params['crop_pos']).astype(int)
bbox_cls = params['bbox_cls']
### generate output bbox
img_size = output_dict['label'].size(1) #shape[1]
context_ratio = np.random.uniform(low=config['min_ctx_ratio'],\
high=config['max_ctx_ratio'])
output_bbox = np.array(get_soft_bbox(input_bbox, img_size, img_size, context_ratio))
mask_in, mask_object_in, mask_context_in = get_masked_image( \
output_dict['label'], input_bbox, bbox_cls)
mask_out, mask_object_out, _ = get_masked_image( \
output_dict['label'], output_bbox)
output_dict['mask_ctx_in'] = mask_context_in.unsqueeze(0) # (1xCxHxW)
output_dict['mask_in'] = mask_in.unsqueeze(0) # (1x1xHxW)
output_dict['mask_out'] = mask_out.unsqueeze(0) # (1x1xHxW)
output_dict['crop_pos'] = torch.from_numpy(crop_pos) # (1x4)
output_dict['label'] = output_dict['label'].unsqueeze(0) # (1x1xHxW)
output_dict['cls'] = torch.LongTensor([bbox_cls])
if transform_img:
output_dict['image'] = output_dict['image'].unsqueeze(0) * \
(1-output_dict['mask_in']).repeat(1,3,1,1)
#else:
# Crop window
x1, y1, x2, y2 = crop_pos # coordinates of crop window
x1 = max(0,x1); y1 = max(0,y1) # make sure in range
width = x2 - x1 + 1; height = y2 - y1 + 1
#
label_crop = label_original[:,:,y1:y2+1,x1:x2+1]
input_bbox_orig = input_bbox.astype(float)
input_bbox_orig = [input_bbox_orig[0] / opt.fineSize * width,
input_bbox_orig[1] / opt.fineSize * height,
input_bbox_orig[2] / opt.fineSize * width,
input_bbox_orig[3] / opt.fineSize * height]
input_bbox_orig = np.array(input_bbox_orig)
output_bbox_orig = output_bbox.astype(float)
output_bbox_orig = [output_bbox_orig[0] / opt.fineSize * width,
output_bbox_orig[1] / opt.fineSize * height,
output_bbox_orig[2] / opt.fineSize * width,
output_bbox_orig[3] / opt.fineSize * height]
output_bbox_orig = np.array(output_bbox_orig)
_, _, mask_ctx_in = get_masked_image(label_crop[0], input_bbox_orig, bbox_cls)
mask_out, _, _ = get_masked_image(label_crop[0], output_bbox_orig)
output_dict['label_orig'] = label_crop
output_dict['mask_ctx_in_orig'] = mask_ctx_in.unsqueeze(0)
output_dict['mask_out_orig'] = mask_out.unsqueeze(0)
output_bbox_global = [x1 + output_bbox_orig[0],
y1 + output_bbox_orig[1],
x1 + output_bbox_orig[2],
y1 + output_bbox_orig[3]]
output_bbox_global = np.array(output_bbox_global)
output_dict['output_bbox'] = torch.from_numpy(output_bbox)
output_dict['output_bbox_global'] = torch.from_numpy(output_bbox_global)
return output_dict