def __getitem__(self, index):
# Sample random scales to use for each image in this batch
item = {'rpn_targets': {}}
target_scale = self.opts[self.cfg_key]['SCALES'][npr.randint(0, high=len(self.opts[self.cfg_key]['SCALES']))]
img = cv2.imread(osp.join(self._data_path, self.annotations[index]['path']))
img_original_shape = img.shape
item['path']= self.annotations[index]['path']
img, im_scale = self._image_resize(img, target_scale, self.opts[self.cfg_key]['MAX_SIZE'])
# restore the [image_height, image_width, scale_factor, max_size]
item['image_info'] = np.array([img.shape[0], img.shape[1], im_scale,
img_original_shape[0], img_original_shape[1]], dtype=np.float)
item['visual'] = Image.fromarray(img)
if self.transform is not None:
item['visual'] = self.transform(item['visual'])
# if self._batch_size > 1:
# # padding the image to MAX_SIZE, so all images can be stacked
# pad_h = self.opts[self.cfg_key]['MAX_SIZE'] - item['visual'].size(1)
# pad_w = self.opts[self.cfg_key]['MAX_SIZE'] - item['visual'].size(2)
# item['visual'] = F.pad(item['visual'], (0, pad_w, 0, pad_h)).data
_annotation = self.annotations[index]
gt_boxes_object = np.zeros((len(_annotation['objects']), 5))
gt_boxes_object[:, 0:4] = np.array([obj['box'] for obj in _annotation['objects']], dtype=np.float) * im_scale
gt_boxes_object[:, 4] = np.array([obj['class'] for obj in _annotation['objects']])
item['objects'] = gt_boxes_object
if self._image_set == 'train': # calculate the RPN target
item['rpn_targets']['object'] = anchor_target_layer(item['visual'], gt_boxes_object, item['image_info'],
self._feat_stride, self._rpn_opts['object'],
mappings = self._rpn_opts['mappings'])
gt_relationships = np.zeros([len(_annotation['objects']), (len(_annotation['objects']))], dtype=np.long)
for rel in _annotation['relationships']:
gt_relationships[rel['sub_id'], rel['obj_id']] = rel['predicate']
item['relations'] = gt_relationships
if self.use_region:
gt_boxes_region = np.zeros((len(_annotation['regions']), self.max_size + 4)) # 4 for box and 40 for sentences
gt_boxes_region[:, 0:4] = np.array([reg['box'] for reg in _annotation['regions']], dtype=np.float) * im_scale
gt_boxes_region[:, 4:] = np.array([np.pad(reg['phrase'],
(0,self.max_size-len(reg['phrase'])),'constant',constant_values=self.voc_sign['end'])
for reg in _annotation['regions']])
item['regions'] = gt_boxes_region
if self._image_set == 'train' and 'region' in self._rpn_opts.keys(): # calculate the RPN target
item['rpn_targets']['region'] = anchor_target_layer(item['visual'], gt_boxes_region, item['image_info'],
self._feat_stride, self._rpn_opts['region'],
mappings = self._rpn_opts['mappings'])
else:
item['regions'] = None
return item
def __getitem__(self, index):
'''
item数据结构:
rpn_targets:
path: 标注所在路径
image_info: 图片信息（image_height, image_width, scale_factor, max_size）
visual: 从array转换为Image的图片
objects: 图片中的object的bbox和class（GroundTruth）
relatonship: 图片中的relationship（GroundTruth）
'''
# Sample random scales to use for each image in this batch
# item是每张图片load进来的数据结构
item = {'rpn_targets': {}}
target_scale = self.opts[self.cfg_key]['SCALES'][npr.randint(
0, high=len(self.opts[self.cfg_key]['SCALES']))]
img = cv2.imread(
osp.join(self._data_path, self.annotations[index]['path']))
img_original_shape = img.shape
item['path'] = self.annotations[index]['path']
# 缩放图片到小于option中的最大值
img, im_scale = self._image_resize(img, target_scale,
self.opts[self.cfg_key]['MAX_SIZE'])
# restore the [image_height, image_width, scale_factor, max_size]
item['image_info'] = np.array([
img.shape[0], img.shape[1], im_scale, img_original_shape[0],
img_original_shape[1]
],
dtype=np.float)
# 将矩阵转换为Image
item['visual'] = Image.fromarray(img)
# 图像归一化？
if self.transform is not None:
item['visual'] = self.transform(item['visual'])
# if self._batch_size > 1:
# # padding the image to MAX_SIZE, so all images can be stacked
# pad_h = self.opts[self.cfg_key]['MAX_SIZE'] - item['visual'].size(1)
# pad_w = self.opts[self.cfg_key]['MAX_SIZE'] - item['visual'].size(2)
# item['visual'] = F.pad(item['visual'], (0, pad_w, 0, pad_h)).data
_annotation = self.annotations[index]
# 从_annotation['objects']中读入object的ground truth（bbox和class）
gt_boxes_object = np.zeros((len(_annotation['objects']), 5))
# box的坐标
gt_boxes_object[:, 0:4] = np.array(
[obj['bbox']
for obj in _annotation['objects']], dtype=np.float) * im_scale
# box对应的class
gt_boxes_object[:, 4] = np.array(
[obj['class'] for obj in _annotation['objects']])
item['objects'] = gt_boxes_object
if self.cfg_key == 'train': # calculate the RPN target
item['rpn_targets']['object'] = anchor_target_layer(
item['visual'],
gt_boxes_object,
item['image_info'],
self._feat_stride,
self._rpn_opts['object'],
mappings=self._rpn_opts['mappings'])
# 获取relationship（predicate）的ground truth
gt_relationships = np.zeros(
[len(_annotation['objects']), (len(_annotation['objects']))],
dtype=np.long)
for rel in _annotation['relationships']:
gt_relationships[rel['sub_id'], rel['obj_id']] = rel['predicate']
item['relations'] = gt_relationships
return item
