def _record_to_tf(self, record):
"""Creates tf.train.SequenceExample object from records.
"""
try:
self._validate_record(record)
except InvalidRecord as e:
# Pop image before displaying record.
record.pop('image_raw')
tf.logging.warning('Invalid record: {} - {}'.format(
e, record
))
return
sequence_vals = {
'label': [],
'xmin': [],
'ymin': [],
'xmax': [],
'ymax': [],
}
for b in record['gt_boxes']:
sequence_vals['label'].append(to_int64(b['label']))
sequence_vals['xmin'].append(to_int64(b['xmin']))
sequence_vals['ymin'].append(to_int64(b['ymin']))
sequence_vals['xmax'].append(to_int64(b['xmax']))
sequence_vals['ymax'].append(to_int64(b['ymax']))
object_feature_lists = {
'label': tf.train.FeatureList(feature=sequence_vals['label']),
'xmin': tf.train.FeatureList(feature=sequence_vals['xmin']),
'ymin': tf.train.FeatureList(feature=sequence_vals['ymin']),
'xmax': tf.train.FeatureList(feature=sequence_vals['xmax']),
'ymax': tf.train.FeatureList(feature=sequence_vals['ymax']),
}
object_features = tf.train.FeatureLists(
feature_list=object_feature_lists
)
feature = {
'width': to_int64(record['width']),
'height': to_int64(record['height']),
'depth': to_int64(record['depth']),
'filename': to_string(record['filename']),
'image_raw': to_bytes(record['image_raw']),
}
# Now build an `Example` protobuf object and save with the writer.
context = tf.train.Features(feature=feature)
example = tf.train.SequenceExample(
feature_lists=object_features, context=context
)
return example
def _record_to_tf(self, record):
"""Creates tf.train.SequenceExample object from records.
"""
try:
self._validate_record(record)
except InvalidRecord as e:
# Pop image before displaying record.
record.pop('image_raw')
tf.logging.warning('Invalid record: {} - {}'.format(
e, record
))
return
sequence_vals = {
'label': [],
'xmin': [],
'ymin': [],
'xmax': [],
'ymax': [],
}
for b in record['gt_boxes']:
sequence_vals['label'].append(to_int64(b['label']))
sequence_vals['xmin'].append(to_int64(b['xmin']))
sequence_vals['ymin'].append(to_int64(b['ymin']))
sequence_vals['xmax'].append(to_int64(b['xmax']))
sequence_vals['ymax'].append(to_int64(b['ymax']))
object_feature_lists = {
'label': tf.train.FeatureList(feature=sequence_vals['label']),
'xmin': tf.train.FeatureList(feature=sequence_vals['xmin']),
'ymin': tf.train.FeatureList(feature=sequence_vals['ymin']),
'xmax': tf.train.FeatureList(feature=sequence_vals['xmax']),
'ymax': tf.train.FeatureList(feature=sequence_vals['ymax']),
}
object_features = tf.train.FeatureLists(
feature_list=object_feature_lists
)
feature = {
'width': to_int64(record['width']),
'height': to_int64(record['height']),
'depth': to_int64(record['depth']),
'filename': to_string(record['filename']),
'image_raw': to_bytes(record['image_raw']),
}
# Now build an `Example` protobuf object and save with the writer.
context = tf.train.Features(feature=feature)
example = tf.train.SequenceExample(
feature_lists=object_features, context=context
)
return example
def _record_to_tf(self, record):
"""Creates tf.train.SequenceExample object from records."""
try:
self._validate_record(record)
except InvalidRecord as e:
# Pop image before displaying record.
record.pop("image_raw")
tf.logging.warning("Invalid record: {} - {}".format(e, record))
return
sequence_vals = {
"label": [],
"xmin": [],
"ymin": [],
"xmax": [],
"ymax": [],
}
for b in record["gt_boxes"]:
sequence_vals["label"].append(to_int64(b["label"]))
sequence_vals["xmin"].append(to_int64(b["xmin"]))
sequence_vals["ymin"].append(to_int64(b["ymin"]))
sequence_vals["xmax"].append(to_int64(b["xmax"]))
sequence_vals["ymax"].append(to_int64(b["ymax"]))
object_feature_lists = {
"label": tf.train.FeatureList(feature=sequence_vals["label"]),
"xmin": tf.train.FeatureList(feature=sequence_vals["xmin"]),
"ymin": tf.train.FeatureList(feature=sequence_vals["ymin"]),
"xmax": tf.train.FeatureList(feature=sequence_vals["xmax"]),
"ymax": tf.train.FeatureList(feature=sequence_vals["ymax"]),
}
object_features = tf.train.FeatureLists(
feature_list=object_feature_lists)
feature = {
"width": to_int64(record["width"]),
"height": to_int64(record["height"]),
"depth": to_int64(record["depth"]),
"filename": to_string(record["filename"]),
"image_raw": to_bytes(record["image_raw"]),
}
# Now build an `Example` protobuf object and save with the writer.
context = tf.train.Features(feature=feature)
example = tf.train.SequenceExample(feature_lists=object_features,
context=context)
return example
def image_to_example(self, classes, image_id):
annotation_path = self.get_image_annotation(image_id)
image_path = self.get_image_path(image_id)
# Read both the image and the annotation into memory.
annotation = read_xml(annotation_path)
image = read_image(image_path)
# TODO: consider alternatives to using Pillow here.
image_pil = Image.open(image_path)
width = image_pil.width
height = image_pil.height
image_pil.close()
obj_vals = {
'label': [],
'xmin': [],
'ymin': [],
'xmax': [],
'ymax': [],
}
objects = annotation.get('object')
if objects is None:
# If there's no bounding boxes, we don't want it
return
for b in annotation['object']:
try:
label_id = classes.index(self._wnids[b['name']])
except ValueError:
continue
(xmin, ymin, xmax, ymax) = adjust_bbox(
xmin=int(b['bndbox']['xmin']), ymin=int(b['bndbox']['ymin']),
xmax=int(b['bndbox']['xmax']), ymax=int(b['bndbox']['ymax']),
old_width=int(annotation['size']['width']),
old_height=int(annotation['size']['height']),
new_width=width, new_height=height
)
obj_vals['label'].append(to_int64(label_id))
obj_vals['xmin'].append(to_int64(xmin))
obj_vals['ymin'].append(to_int64(ymin))
obj_vals['xmax'].append(to_int64(xmax))
obj_vals['ymax'].append(to_int64(ymax))
if len(obj_vals['label']) == 0:
# No bounding box matches the available classes.
return
object_feature_lists = {
'label': tf.train.FeatureList(feature=obj_vals['label']),
'xmin': tf.train.FeatureList(feature=obj_vals['xmin']),
'ymin': tf.train.FeatureList(feature=obj_vals['ymin']),
'xmax': tf.train.FeatureList(feature=obj_vals['xmax']),
'ymax': tf.train.FeatureList(feature=obj_vals['ymax']),
}
object_features = tf.train.FeatureLists(
feature_list=object_feature_lists
)
sample = {
'width': to_int64(width),
'height': to_int64(height),
'depth': to_int64(3),
'filename': to_string(annotation['filename']),
'image_raw': to_bytes(image),
}
# Now build an `Example` protobuf object and save with the writer.
context = tf.train.Features(feature=sample)
example = tf.train.SequenceExample(
feature_lists=object_features, context=context
)
return example
def image_to_example(self, classes, image_id):
annotation_path = self.get_image_annotation(image_id)
image_path = self.get_image_path(image_id)
# Read both the image and the annotation into memory.
annotation = read_xml(annotation_path)
image = read_image(image_path)
object_features_values = {
'label': [],
'xmin': [],
'ymin': [],
'xmax': [],
'ymax': [],
}
for b in annotation['object']:
try:
label_id = classes.index(b['name'])
except ValueError:
continue
object_features_values['label'].append(to_int64(label_id))
object_features_values['xmin'].append(to_int64(
b['bndbox']['xmin']))
object_features_values['ymin'].append(to_int64(
b['bndbox']['ymin']))
object_features_values['xmax'].append(to_int64(
b['bndbox']['xmax']))
object_features_values['ymax'].append(to_int64(
b['bndbox']['ymax']))
if len(object_features_values['label']) == 0:
# No bounding box matches the available classes.
return
object_feature_lists = {
'label':
tf.train.FeatureList(feature=object_features_values['label']),
'xmin':
tf.train.FeatureList(feature=object_features_values['xmin']),
'ymin':
tf.train.FeatureList(feature=object_features_values['ymin']),
'xmax':
tf.train.FeatureList(feature=object_features_values['xmax']),
'ymax':
tf.train.FeatureList(feature=object_features_values['ymax']),
}
object_features = tf.train.FeatureLists(
feature_list=object_feature_lists)
sample = {
'width': to_int64(int(annotation['size']['width'])),
'height': to_int64(int(annotation['size']['height'])),
'depth': to_int64(int(annotation['size']['depth'])),
'filename': to_string(annotation['filename']),
'image_raw': to_bytes(image),
}
# Now build an `Example` protobuf object and save with the writer.
context = tf.train.Features(feature=sample)
example = tf.train.SequenceExample(feature_lists=object_features,
context=context)
return example