def _generate_examples()

in tensorflow_datasets/vision_language/refcoco/refcoco.py [0:0]

• 41 lines of code
• 15 McCabe index (conditional complexity)

def _generate_examples(refcoco_json, dataset, dataset_partition, split):
  """Generates examples of images and its refexps & ground truth bboxes.

  Args:
    refcoco_json: contents of the annotation file.
    dataset: str specifying the dataset (refcoco, refcoco+, refcocog)
    dataset_partition: str specifying the partition for the dataset
    split: str specifying the split of the dataset_partition

  Yields:
    image_id and example tuple
  """
  refcoco_anns = refcoco_json['ref']
  coco_anns = refcoco_json['coco_anns']

  # Collect all referring expressions for a given image.
  imageid2annref = collections.defaultdict(list)
  for r in refcoco_anns:
    if r['dataset'] == dataset and r[
        'dataset_partition'] == dataset_partition and r['split'] == split:
      imageid2annref[r['image_id']].append(r)

  # Process all the referring expressions and ground truth annotations for
  # a given COCO image.
  for image_id in sorted(imageid2annref.keys()):
    coco_image = coco_anns[str(image_id)]
    image_info = coco_image['info']
    example = {
        'image_filename': image_info['file_name'],
        'image/id': image_id,
        'coco_annotations': [],
        'objects': [],
    }

    # Collect ground truth bboxes.
    for ann in sorted(coco_image['anns'], key=operator.itemgetter('id')):
      example['coco_annotations'].append(_extract_annotation(ann, image_info))

    # Collect referring expressions.
    for r in sorted(imageid2annref[image_id],
                    key=operator.itemgetter('ref_id')):
      obj = _extract_annotation(r['ann'], image_info)

      refexp = []
      for s in sorted(r['sentences'], key=operator.itemgetter('sent_id')):
        refexp.append({
            'raw': s['raw'],
            'refexp_id': s['sent_id'],
        })

      # Match the referring expression to its corresponding bbox in the ground
      # truth list.
      gt_box_index = [
          i for i, v in enumerate(example['coco_annotations'])
          if v['id'] == r['ann']['id']
      ]
      if len(gt_box_index) != 1:
        raise ValueError(f'gt_box_index does not have length 1: {gt_box_index}')
      gt_box_index = gt_box_index[0]

      obj.update({
          'refexp': refexp,
          'gt_box_index': gt_box_index,
      })
      example['objects'].append(obj)

    yield image_id, example