def _imagePreprocess(self, blob, fix_size=True):
keep = np.arange(blob['gt_boxes'].shape[0])
if self.augmentation:
if self.augImageOnly is not None:
blob['data'] = self.augImageOnly(blob['data'])
if self.augObjdet is not None: blob['data'], blob['gt_boxes'], _, keep, _ = \
self.augObjdet(image=blob['data'], boxes=blob['gt_boxes'], boxes_keep=keep)
# choose one predefined size, TODO: support multi-instance batch
random_scale_ind = np.random.randint(0, high=len(cfg.SCALES))
blob['data'], im_scale = prep_im_for_blob(blob['data'],
cfg.SCALES[random_scale_ind],
cfg.TRAIN.COMMON.MAX_SIZE,
fix_size)
# modify bounding boxes according to resize parameters
blob['im_info'][:2] = (blob['data'].shape[0], blob['data'].shape[1])
blob['im_info'][2:4] = (im_scale['y'], im_scale['x'])
blob['gt_boxes'][:, :-1][:, 0::2] *= im_scale['x']
blob['gt_boxes'][:, :-1][:, 1::2] *= im_scale['y']
blob['data'] = image_normalize(blob['data'],
mean=self.pixel_means,
std=self.pixel_stds)
blob['node_inds'] = blob['node_inds'][keep]
blob['parent_lists'] = [
blob['parent_lists'][p_ind] for p_ind in list(keep)
]
blob['child_lists'] = [
blob['child_lists'][c_ind] for c_ind in list(keep)
]
return blob
def _imagePreprocess(self, blob, fix_size=False):
assert not fix_size, "When grasp labels are included, the input image can not be fixed-size."
keep_b = np.arange(blob['gt_boxes'].shape[0])
keep_g = np.arange(blob['gt_grasps'].shape[0])
if self.augmentation:
blob['data'] = self.augImageOnly(blob['data'])
blob['data'], blob['gt_boxes'], blob['gt_grasps'], keep_b, keep_g = \
self.augObjdet(image=blob['data'], boxes=blob['gt_boxes'], grasps=blob['gt_grasps'],
boxes_keep=keep_b, grasps_keep=keep_g)
# choose one predefined size, TODO: support multi-instance batch
random_scale_ind = np.random.randint(0, high=len(cfg.SCALES))
blob['data'], im_scale = prep_im_for_blob(blob['data'],
cfg.SCALES[random_scale_ind],
cfg.TRAIN.COMMON.MAX_SIZE,
fix_size)
blob['im_info'][:2] = (blob['data'].shape[0], blob['data'].shape[1])
blob['im_info'][2:4] = (im_scale['y'], im_scale['x'])
# modify bounding boxes according to resize parameters
blob['gt_boxes'][:, :-1][:, 0::2] *= im_scale['x']
blob['gt_boxes'][:, :-1][:, 1::2] *= im_scale['y']
blob['gt_grasps'][:, 0::2] *= im_scale['x']
blob['gt_grasps'][:, 1::2] *= im_scale['y']
blob['gt_grasp_inds'] = blob['gt_grasp_inds'][keep_g]
blob['data'] = image_normalize(blob['data'],
mean=self.pixel_means,
std=self.pixel_stds)
blob['node_inds'] = blob['node_inds'][keep_b]
blob['parent_lists'] = [
blob['parent_lists'][p_ind] for p_ind in list(keep_b)
]
blob['child_lists'] = [
blob['child_lists'][c_ind] for c_ind in list(keep_b)
]
return blob
def _imagePreprocess(self, blob, fix_size = True):
keep = np.arange(blob['gt_grasps'].shape[0])
if self.augmentation:
blob['data'] = self.augImageOnly(blob['data'])
blob['data'], _, blob['gt_grasps'], _, _ = self.augmGraspdet(image=blob['data'], grasps=blob['gt_grasps'], grasps_keep=keep)
# choose one predefined size, TODO: support multi-instance batch
random_scale_ind = np.random.randint(0, high=len(cfg.SCALES))
blob['data'], im_scale = prep_im_for_blob(blob['data'], cfg.SCALES[random_scale_ind], cfg.TRAIN.COMMON.MAX_SIZE, fix_size)
blob['im_info'][:2] = (blob['data'].shape[0], blob['data'].shape[1])
blob['im_info'][2:4] = (im_scale['y'], im_scale['x'])
blob['gt_grasps'][:, 0::2] *= im_scale['x']
blob['gt_grasps'][:, 1::2] *= im_scale['y']
blob['data'] = image_normalize(blob['data'], mean=cfg.PIXEL_MEANS, std=cfg.PIXEL_STDS)
return blob