datasetinsights.datasets¶

datasetinsights.datasets.base¶

class datasetinsights.datasets.base.Dataset¶

Bases: object

Abstract base class for datasets.

static create(name, **kwargs)¶

Create a new instance of the dataset subclass

Parameters

name (str) – unique identifier for a dataset subclass
config (dict) – parameters specific to each dataset subclass used to create a dataset instance

Returns

an instance of the specified dataset subclass

static find(name)¶

Find Dataset subclass based on the given name

Parameters: name (str) – unique identifier for a dataset subclass
Returns: a label of the specified dataset subclass

datasetinsights.datasets.cityscapes¶

class datasetinsights.datasets.cityscapes.Cityscapes(*, data_root='/data', split='train', transforms=None, **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

Parameters

data_root (str) – root directory prefix of all datasets
split (str) – indicate split type of the dataset (train|val|test)

root¶

root directory of the dataset

Type: str

split¶

indicate split type of the dataset (train|val|test)

Type: str

download(cloud_path)¶

Download cityscapes dataset

Note

The current implementation assumes a GCS cloud path. Should we keep this method here if we want to support other cloud storage system?

Parameters: cloud_path (str) – cloud path of the dataset

datasetinsights.datasets.coco¶

class datasetinsights.datasets.coco.CocoDetection(*, data_root='/data', split='train', transforms=None, remove_examples_without_boxes=True, **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

http://cocodataset.org/#detection-2019

download(cloud_path='data/coco')¶

datasetinsights.datasets.coco.convert_coco2canonical(coco_annotation)¶

convert from a tuple of image and coco style dictionary describing the bboxes to a tuple of image, List of BBox2D :param coco_annotation: image and coco style dictionary :type coco_annotation: tuple

Returns: a tuple of image, List of BBox2D

datasetinsights.datasets.exceptions¶

exception datasetinsights.datasets.exceptions.DatasetNotFoundError¶

Bases: Exception

Raise when a dataset file can’t be found.

datasetinsights.datasets.groceries_real¶

class datasetinsights.datasets.groceries_real.GoogleGroceriesReal(data_root, split='train', transforms=None)¶

Bases: datasetinsights.datasets.base.Dataset

Google GroceriesReal Dataset

This dataset include groceries images and annotations in this paper. Please reach out to the authors to request the dataset.

The dataset should contain the following structures:

Train Data/
file0.tfexample file1.tfexample …
Eval Data/
file234.tfexample …
Label Map/label_map_with_path_64_retail.txt

root¶

Local directory of the dataset is.

Type: str

split¶

Indicate split type of the dataset. Allowed value are (train|test).

Type: str

transforms¶: Callable transformation

label_mappings¶

a dict of {label_id: label_name} mapping loaded from label_map_with_path_64_retail.txt

Type: dict

GCS_PATH = 'data/google_groceries'¶

LABEL_FILE = 'Label Map/label_map_with_path_64_retail.txt'¶

LABEL_ZIP = 'Label Map.zip'¶

LOCAL_PATH = 'google_groceries'¶

SPLITS_FOLDER = {'test': 'Eval Data', 'train': 'Train Data'}¶

SPLITS_ZIP = {'test': 'Eval Data.zip', 'train': 'Train Data.zip'}¶

TF_FEATURES = {'image/channels': FixedLenFeature(shape=[], dtype=tf.int64, default_value=None), 'image/encoded': FixedLenFeature(shape=[], dtype=tf.string, default_value=None), 'image/format': FixedLenFeature(shape=[], dtype=tf.string, default_value=None), 'image/height': FixedLenFeature(shape=[], dtype=tf.int64, default_value=None), 'image/object/bbox/occluded': FixedLenFeature(shape=[], dtype=tf.float32, default_value=None), 'image/object/bbox/xmax': VarLenFeature(dtype=tf.float32), 'image/object/bbox/xmin': VarLenFeature(dtype=tf.float32), 'image/object/bbox/ymax': VarLenFeature(dtype=tf.float32), 'image/object/bbox/ymin': VarLenFeature(dtype=tf.float32), 'image/object/class/label': VarLenFeature(dtype=tf.int64), 'image/object/class/text': VarLenFeature(dtype=tf.string), 'image/width': FixedLenFeature(shape=[], dtype=tf.int64, default_value=None)}¶

download()¶: Download dataset from GCS

class datasetinsights.datasets.groceries_real.GroceriesReal(*, data_path='/data', split='train', transforms=None, version='v3', **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

Unity’s Groceries Real Dataset.

During the class instantiation, it would check whether the data is downloaded or not. If there is no dataset, it would raise an error. Please make sure you download the dataset before use this class.

The dataset contain the following structures:

images/
- IMG_4185.JPG
- IMG_4159.JPG
- …
annotation_definition.json file stores label_id to label_name mappings
annotations.json file stores 2d bounding box records:

groceries_real_<split>.txt file which stores space deplimted list

of the indices in the annotations.json file, for a given data split.

dataset_directory¶

Local directory where groceries_real dataset is saved.

Type: str

split¶

Indicate split type of the dataset. Allowed value are (train|val|test|test_high_ratio|test_low_ratio) test_high_ratio: Split of test dataset with high foreground-to-background ratio. test_low_ratio: Split of test dataset with low foreground-to-background ratio.

Type: str

transforms¶: Callable transformation

annotations¶

A list of dict that contain image file path and 2D bounding box annotations.

Type: list

split_indices¶

A list of indices for this dataset split.

Type: list

label_mappings¶

a dict of {label_id: label_name} mapping

Type: dict

version¶

version of GroceriesReal dataset, e.g. “v3”.

Type: str

default version="v3".

ANNOTATION_FILE = 'annotations.json'¶

ARCHIVE_FILES = {'v3': 'v3.zip'}¶

DEFINITION_FILE = 'annotation_definitions.json'¶

SPLITS = {'test': 'groceries_real_test.txt', 'test_high_contrast': 'groceries_real_test_high_contrast.txt', 'test_high_ratio': 'groceries_real_test_high_ratio.txt', 'test_low_contrast': 'groceries_real_test_low_contrast.txt', 'test_low_ratio': 'groceries_real_test_low_ratio.txt', 'train': 'groceries_real_train.txt', 'train_medium': 'groceries_real_train_medium.txt', 'train_small': 'groceries_real_train_small.txt', 'val': 'groceries_real_val.txt'}¶

SUBFOLDER = 'groceries'¶

static is_dataset_files_present(data_path)¶

datasetinsights.datasets.kitti¶

class datasetinsights.datasets.kitti.Kitti(root='/data/kitti', split='train', indices_file: str = None)¶

Bases: datasetinsights.datasets.base.Dataset

dataloader for kitti dataset http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d

download(cloud_path='data/kitti')¶

download_kitti_zips(cloud_client, cloud_path='data/kitti')¶

class datasetinsights.datasets.kitti.KittiBox3d(*, label: str, position, dimensions, angle, sample_idx, score: float = 1.0)¶

Bases: object

class to represent a bounding box for the kitti dataset. Note that this style of bounding box is not primarily supported. The canonical 3d bounding box class for this repo is the class BBox3D. Reference code for KittiBox3d found at http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d https://github.com/bostondiditeam/kitti/tree/master/resources/devkit_object To convert from kitti style to nuscenes use the method convert_kitti2nu.

class datasetinsights.datasets.kitti.KittiTransforms(calib_filename)¶

Bases: object

Class to hold transformation matrices for a kitti data sample see more at https://github.com/yanii/kitti-pcl/blob/master/KITTI_README.TXT

property projection_mat_left¶

Projection matrix for left image (to project bounding box coordinate to image coordinates).

Type: Returns

property projection_rect_combined¶

Merge rectification and projection into one matrix.

Returns: combined rectification and projection matrix

property r0_rect¶

Quaternion to rectify camera frame.

Type: Returns

property velo_to_cam_rotation¶

Rotation matrix takes a point in Velodyne coordinates and transforms it: into the

coordinate system of the left video camera. Likewise it serves as a representation of the Velodyne coordinate frame in camera coordinates. Returns: Rotation matrix

property velo_to_cam_translation¶

Translation matrix takes a point in Velodyne coordinates and transforms: it into the

coordinate system of the left video camera. Likewise it serves as a representation of the Velodyne coordinate frame in camera coordinates. Returns: Translation matrix

datasetinsights.datasets.kitti.convert_kitti2nu(*, bbox: datasetinsights.datasets.kitti.KittiBox3d, transforms: datasetinsights.datasets.kitti.KittiTransforms) → datasetinsights.io.bbox.BBox3D ¶

convert a bounding box from kitti format to nuscenes format

Parameters

bbox – bounding box in kitti format
transforms – camera transforms

Returns:

datasetinsights.datasets.kitti.read_kitti_calib(filename)¶: Read the camera 2 calibration matrix from box text file

datasetinsights.datasets.kitti.read_kitti_objects(filename)¶

datasetinsights.datasets.nuscenes¶

class datasetinsights.datasets.nuscenes.NuscenesDataLoader(*, data_root='/data', split='train', version='v1.0', full_dataset: bool = False, download=False, sensor_data: List[str] = ['CAM_FRONT', 'CAM_FRONT_RIGHT', 'CAM_BACK_RIGHT', 'CAM_BACK', 'CAM_BACK_LEFT', 'CAM_FRONT_LEFT', 'RADAR_FRONT', 'RADAR_FRONT_LEFT', 'RADAR_FRONT_RIGHT', 'RADAR_BACK_LEFT', 'RADAR_BACK_RIGHT', 'LIDAR_TOP'], coordinates='global')¶

Bases: torch.utils.data.dataset.Dataset

Nuscenes Dataset https://www.nuscenes.org/ :param data_root: root directory prefix of all datasets :type data_root: str :param split: test or valuation. see SPLITS constant for possible values :param version: numerical version of nuscenes dataset. :param sensor_data: what sensor data to use, can be any of the strings in :param SENSOR_DATA_KEYS: :param full_dataset: whether to use the full dataset. If false, use the mini :param dataset: :param download: whether or no to download the nuscenes dataset from gcs. If :param false: :param it is assumed the: :param dataset is stored at the location specified by root: :param coordinates: whether to use coordinates in the local coordinate system :param of the sensor or the global: :param log coordinates. Options: ‘sensor’ or ‘global’

root¶: directory where nuscenes data is stored. If no value is provided, the specified version of nuscenes is downloaded.

download(version='v1.0-mini')¶: download the nuscenes dataset version specified to /data/sets/nuscenes

datasetinsights.datasets.nyudepth¶

Load NYU Depth V2 data.

Downloaded NYUDepth dataset link:: https://drive.google.com/drive/folders/1TzwfNA5JRFTPO-kHMU___kILmOEodoBo

This dataset is not the official NYU Depth V2 datset. The data are preprocessed. For more information, see the link below:

https://github.com/ialhashim/DenseDepth

class datasetinsights.datasets.nyudepth.NyuDepth(data_root='/data', transforms=None, split='train', cloud_path='data/nyudepth')¶

Bases: datasetinsights.datasets.base.Dataset

root¶

local directory where nyu_v2 dataset is saved

Type: str

nyu2_data¶

training or testing data

Type: Pandas DataFrame

split¶

test or valuation. see SPLITS constant for possible values

Type: str

download(cloud_path)¶

Download nyu_v2 dataset The directory structure of the downloaded data is |–self.root

|–nyudepth
|–nyu_data.zip |–data

|–nyu2_test.csv |–nyu2_test

|–00000_colors.png |–00000_depth.png … |–01448_colors.png |–01448_depth.png

|–nyu2_train.csv |–nyu2_train

|–basement_0001a_out
|–1.jpg |–1.png … |–281.jpg |–281.png

… |–study_room_0005b_out

|–1.jpg |–1.png … |–133.jpg |–133.png

Parameters: cloud_path (str) – cloud path of the dataset

datasetinsights.datasets.synthetic¶

Simulation Dataset Catalog

class datasetinsights.datasets.synthetic.SynDetection2D(*, data_path=None, split='all', transforms=None, version='0.0.1', def_id=4, train_split_ratio=0.9, random_seed=47, **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

Synthetic dataset for 2D object detection.

During the class instantiation, it would check whether the data files such as annotations.json, images.png are present, if not it’ll check whether a compressed dataset file is present which contains the necessary files, if not it’ll raise an error.

See synthetic dataset schema documentation for more details. <https://datasetinsights.readthedocs.io/en/latest/Synthetic_Dataset_Schema.html>

catalog¶

catalog of all captures in this dataset

Type: list

transforms¶: callable transformation that applies to a pair of capture, annotation. Capture is the information captured by the sensor, in this case an image, and annotations, which in this dataset are 2d bounding box coordinates and labels.

split¶

indicate split type of the dataset (train|val|test)

Type: str

label_mappings¶

a dict of {label_id: label_name} mapping

Type: dict

ARCHIVE_FILE = 'SynthDet.zip'¶

SUBFOLDER = 'synthetic'¶

static is_dataset_files_present(data_path)¶

datasetinsights.datasets.synthetic.read_bounding_box_2d(annotation, label_mappings=None)¶

Convert dictionary representations of 2d bounding boxes into objects of the BBox2D class

Parameters

annotation (List[dict]) – 2D bounding box annotation
label_mappings (dict) – a dict of {label_id: label_name} mapping

Returns

A list of 2D bounding box objects

datasetinsights.datasets.synthetic.read_bounding_box_3d(annotation, label_mappings=None)¶

Convert dictionary representations of 3d bounding boxes into objects of the BBox3d class

Parameters

annotation (List[dict]) – 3D bounding box annotation
label_mappings (dict) – a dict of {label_id: label_name} mapping

Returns

A list of 3d bounding box objects

class datasetinsights.datasets.Cityscapes(*, data_root='/data', split='train', transforms=None, **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

Parameters

data_root (str) – root directory prefix of all datasets
split (str) – indicate split type of the dataset (train|val|test)

root¶

root directory of the dataset

Type: str

split¶

indicate split type of the dataset (train|val|test)

Type: str

download(cloud_path)¶

Download cityscapes dataset

Note

The current implementation assumes a GCS cloud path. Should we keep this method here if we want to support other cloud storage system?

Parameters: cloud_path (str) – cloud path of the dataset

class datasetinsights.datasets.CocoDetection(*, data_root='/data', split='train', transforms=None, remove_examples_without_boxes=True, **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

http://cocodataset.org/#detection-2019

download(cloud_path='data/coco')¶

class datasetinsights.datasets.Dataset¶

Bases: object

Abstract base class for datasets.

static create(name, **kwargs)¶

Create a new instance of the dataset subclass

Parameters

name (str) – unique identifier for a dataset subclass
config (dict) – parameters specific to each dataset subclass used to create a dataset instance

Returns

an instance of the specified dataset subclass

static find(name)¶

Find Dataset subclass based on the given name

Parameters: name (str) – unique identifier for a dataset subclass
Returns: a label of the specified dataset subclass

class datasetinsights.datasets.GoogleGroceriesReal(data_root, split='train', transforms=None)¶

Bases: datasetinsights.datasets.base.Dataset

Google GroceriesReal Dataset

This dataset include groceries images and annotations in this paper. Please reach out to the authors to request the dataset.

The dataset should contain the following structures:

Train Data/
file0.tfexample file1.tfexample …
Eval Data/
file234.tfexample …
Label Map/label_map_with_path_64_retail.txt

root¶

Local directory of the dataset is.

Type: str

split¶

Indicate split type of the dataset. Allowed value are (train|test).

Type: str

transforms¶: Callable transformation

label_mappings¶

a dict of {label_id: label_name} mapping loaded from label_map_with_path_64_retail.txt

Type: dict

GCS_PATH = 'data/google_groceries'¶

LABEL_FILE = 'Label Map/label_map_with_path_64_retail.txt'¶

LABEL_ZIP = 'Label Map.zip'¶

LOCAL_PATH = 'google_groceries'¶

SPLITS_FOLDER = {'test': 'Eval Data', 'train': 'Train Data'}¶

SPLITS_ZIP = {'test': 'Eval Data.zip', 'train': 'Train Data.zip'}¶

TF_FEATURES = {'image/channels': FixedLenFeature(shape=[], dtype=tf.int64, default_value=None), 'image/encoded': FixedLenFeature(shape=[], dtype=tf.string, default_value=None), 'image/format': FixedLenFeature(shape=[], dtype=tf.string, default_value=None), 'image/height': FixedLenFeature(shape=[], dtype=tf.int64, default_value=None), 'image/object/bbox/occluded': FixedLenFeature(shape=[], dtype=tf.float32, default_value=None), 'image/object/bbox/xmax': VarLenFeature(dtype=tf.float32), 'image/object/bbox/xmin': VarLenFeature(dtype=tf.float32), 'image/object/bbox/ymax': VarLenFeature(dtype=tf.float32), 'image/object/bbox/ymin': VarLenFeature(dtype=tf.float32), 'image/object/class/label': VarLenFeature(dtype=tf.int64), 'image/object/class/text': VarLenFeature(dtype=tf.string), 'image/width': FixedLenFeature(shape=[], dtype=tf.int64, default_value=None)}¶

download()¶: Download dataset from GCS

class datasetinsights.datasets.GroceriesReal(*, data_path='/data', split='train', transforms=None, version='v3', **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

Unity’s Groceries Real Dataset.

During the class instantiation, it would check whether the data is downloaded or not. If there is no dataset, it would raise an error. Please make sure you download the dataset before use this class.

The dataset contain the following structures:

images/
- IMG_4185.JPG
- IMG_4159.JPG
- …
annotation_definition.json file stores label_id to label_name mappings
annotations.json file stores 2d bounding box records:

groceries_real_<split>.txt file which stores space deplimted list

of the indices in the annotations.json file, for a given data split.

dataset_directory¶

Local directory where groceries_real dataset is saved.

Type: str

split¶

Indicate split type of the dataset. Allowed value are (train|val|test|test_high_ratio|test_low_ratio) test_high_ratio: Split of test dataset with high foreground-to-background ratio. test_low_ratio: Split of test dataset with low foreground-to-background ratio.

Type: str

transforms¶: Callable transformation

annotations¶

A list of dict that contain image file path and 2D bounding box annotations.

Type: list

split_indices¶

A list of indices for this dataset split.

Type: list

label_mappings¶

a dict of {label_id: label_name} mapping

Type: dict

version¶

version of GroceriesReal dataset, e.g. “v3”.

Type: str

default version="v3".

ANNOTATION_FILE = 'annotations.json'¶

ARCHIVE_FILES = {'v3': 'v3.zip'}¶

DEFINITION_FILE = 'annotation_definitions.json'¶

SPLITS = {'test': 'groceries_real_test.txt', 'test_high_contrast': 'groceries_real_test_high_contrast.txt', 'test_high_ratio': 'groceries_real_test_high_ratio.txt', 'test_low_contrast': 'groceries_real_test_low_contrast.txt', 'test_low_ratio': 'groceries_real_test_low_ratio.txt', 'train': 'groceries_real_train.txt', 'train_medium': 'groceries_real_train_medium.txt', 'train_small': 'groceries_real_train_small.txt', 'val': 'groceries_real_val.txt'}¶

SUBFOLDER = 'groceries'¶

static is_dataset_files_present(data_path)¶

class datasetinsights.datasets.NyuDepth(data_root='/data', transforms=None, split='train', cloud_path='data/nyudepth')¶

Bases: datasetinsights.datasets.base.Dataset

root¶

local directory where nyu_v2 dataset is saved

Type: str

nyu2_data¶

training or testing data

Type: Pandas DataFrame

split¶

test or valuation. see SPLITS constant for possible values

Type: str

download(cloud_path)¶

Download nyu_v2 dataset The directory structure of the downloaded data is |–self.root

|–nyudepth
|–nyu_data.zip |–data

|–nyu2_test.csv |–nyu2_test

|–00000_colors.png |–00000_depth.png … |–01448_colors.png |–01448_depth.png

|–nyu2_train.csv |–nyu2_train

|–basement_0001a_out
|–1.jpg |–1.png … |–281.jpg |–281.png

… |–study_room_0005b_out

|–1.jpg |–1.png … |–133.jpg |–133.png

Parameters: cloud_path (str) – cloud path of the dataset

class datasetinsights.datasets.SynDetection2D(*, data_path=None, split='all', transforms=None, version='0.0.1', def_id=4, train_split_ratio=0.9, random_seed=47, **kwargs)¶

Bases: datasetinsights.datasets.base.Dataset

Synthetic dataset for 2D object detection.

During the class instantiation, it would check whether the data files such as annotations.json, images.png are present, if not it’ll check whether a compressed dataset file is present which contains the necessary files, if not it’ll raise an error.

See synthetic dataset schema documentation for more details. <https://datasetinsights.readthedocs.io/en/latest/Synthetic_Dataset_Schema.html>

catalog¶

catalog of all captures in this dataset

Type: list

transforms¶: callable transformation that applies to a pair of capture, annotation. Capture is the information captured by the sensor, in this case an image, and annotations, which in this dataset are 2d bounding box coordinates and labels.

split¶

indicate split type of the dataset (train|val|test)

Type: str

label_mappings¶

a dict of {label_id: label_name} mapping

Type: dict

ARCHIVE_FILE = 'SynthDet.zip'¶

SUBFOLDER = 'synthetic'¶

static is_dataset_files_present(data_path)¶