def parse_sample(sample):
video_path, video_metadata = sample
return fos.Sample(
filepath=video_path,
metadata=video_metadata,
tags=tags,
)
def parse_sample(sample):
image_path, image_metadata = sample
return fos.Sample(
filepath=image_path,
metadata=image_metadata,
tags=tags,
)
def parse_sample(sample):
video_path, video_metadata, frames = sample
sample = fos.Sample(
filepath=video_path, metadata=video_metadata, tags=tags,
)
if frames is not None:
sample.frames.update(frames)
return sample
def parse_sample(sample):
image_path, image_metadata, label = sample
sample = fos.Sample(
filepath=image_path, metadata=image_metadata, tags=tags,
)
if label is not None:
sample[label_field] = label
return sample
def parse_sample(sample):
sample_parser.with_sample(sample)
image_path = sample_parser.get_image_path()
if sample_parser.has_image_metadata:
metadata = sample_parser.get_image_metadata()
else:
metadata = None
return fos.Sample(filepath=image_path, metadata=metadata, tags=tags)
def parse_sample(sample):
image_path, image_metadata, label = sample
sample = fos.Sample(
filepath=image_path, metadata=image_metadata, tags=tags,
)
if isinstance(label, dict):
sample.update_fields(label)
elif label is not None:
sample[label_field] = label
return sample
def parse_sample(sample):
sample_parser.with_sample(sample)
video_path = sample_parser.get_video_path()
if sample_parser.has_video_metadata:
metadata = sample_parser.get_video_metadata()
else:
metadata = None
sample = fos.Sample(filepath=video_path, metadata=metadata, tags=tags)
frames = sample_parser.get_frame_labels()
if frames is not None:
sample.frames.update(frames)
return sample
def parse_sample(sample):
sample_parser.with_sample(sample)
image_path = sample_parser.get_image_path()
if sample_parser.has_image_metadata:
metadata = sample_parser.get_image_metadata()
else:
metadata = None
label = sample_parser.get_label()
sample = fos.Sample(filepath=image_path, metadata=metadata, tags=tags)
if isinstance(label, dict):
sample.update_fields(label)
elif label is not None:
sample[label_field] = label
return sample
def load_open_images_dataset(
dataset_name,
images_dir,
bounding_boxes_path=None,
image_labels_path=None,
predictions_path=None,
prediction_field_name="predicted_detections",
class_descriptions_path=None,
load_images_with_preds=False,
max_num_images=-1,
):
"""Loads an Open Images format dataset into FiftyOne.
**Note** If this takes a long time it is highly recommended to save the
dataset via:
dataset.persistent = True
such that this function only needs to be called once!
Args:
dataset_name: the name of the dataset to create in FiftyOne.
images_dir: directory where images are stored. Images should be in
<open-images-id>.jpg format
bounding_boxes_path: path to the expanded-hierarchy annotation bounding
boxes CSV
image_labels_path: path to the expanded-hierarchy annotation image
labels CSV
predictions_path: path to the predicted bounding boxes CSV
prediction_field_name: the name of the field to save the predictions
under. Useful if other predictions may be added later
class_descriptions_path: optional metadata file. if provided, the
MID labels are mapped to descriptive labels
load_images_with_preds: if True, skip any images that do not have
predictions
Returns:
a :class:`fiftyone.core.dataset.Dataset` instance
"""
# pylint: disable=unsubscriptable-object
# read data from disk
all_location_annotations = (pd.read_csv(bounding_boxes_path)
if bounding_boxes_path else None)
all_label_annotations = (pd.read_csv(image_labels_path)
if image_labels_path else None)
if predictions_path:
all_predictions = pd.read_csv(predictions_path)
all_predictions.rename(columns={"Score": "Confidence"}, inplace=True)
else:
all_predictions = None
class_descriptions = (pd.read_csv(
class_descriptions_path, header=None, index_col=0)
if class_descriptions_path else None)
# map label MID to descriptive label
if class_descriptions is not None:
for df in [
all_location_annotations,
all_label_annotations,
all_predictions,
]:
if df is None:
continue
temp = class_descriptions.loc[df["LabelName"], 1]
temp.index = df.index
df["LabelName"] = temp
if load_images_with_preds:
img_paths = [
os.path.join(images_dir, image_id + ".jpg")
for image_id in set(all_predictions["ImageID"])
]
else:
img_paths = glob.glob(os.path.join(images_dir, "*.jpg"))
if max_num_images != -1:
img_paths = img_paths[:max_num_images]
print("Parsing CSV labels...")
_samples = []
with fou.ProgressBar(img_paths) as pb:
for image_path in pb(img_paths):
image_id = os.path.splitext(os.path.basename(image_path))[0]
kwargs = {"filepath": image_path, OPEN_IMAGES_ID: image_id}
# parse ground truth image labels
if all_label_annotations is not None:
cur_lab_anns = all_label_annotations.query("ImageID == '%s'" %
image_id)
if not cur_lab_anns.empty:
kwargs[GT_IMAGE_LABELS] = df2classifications(cur_lab_anns)
# parse ground truth bounding boxes
if all_location_annotations is not None:
cur_loc_anns = all_location_annotations.query(
"ImageID == '%s'" % image_id)
if not cur_loc_anns.empty:
kwargs[GT_DETECTIONS] = df2detections(cur_loc_anns)
# parse prediction bounding boxes
if all_predictions is not None:
cur_preds = all_predictions.query("ImageID == '%s'" % image_id)
if not cur_preds.empty:
kwargs[prediction_field_name] = df2detections(cur_preds)
_samples.append(fos.Sample(**kwargs))
print("Creating FiftyOne Dataset...")
dataset = fod.Dataset(dataset_name)
dataset.add_samples(_samples)
return dataset
