ObjectDetectionLabels

class ObjectDetectionLabels

Bases: Labels

A set of boxes and associated class_ids and scores.

Implemented using the Tensorflow Object Detection API’s BoxList class.

__init__(npboxes: array, class_ids: array, scores: Optional[array] = None)

Construct a set of object detection labels.

Parameters

• npboxes (array) – float numpy array of size nx4 with cols ymin, xmin, ymax, xmax. Should be in pixel coordinates within the global frame of reference.

• class_ids (array) – int numpy array of size n with class ids

• scores (Optional[array]) – float numpy array of size n

Methods

__init__(npboxes, class_ids[, scores])	Construct a set of object detection labels.
assert_equal(expected_labels)
concatenate(labels1, labels2)	Return concatenation of labels.
filter_by_aoi(aoi_polygons)	Return a copy of these labels filtered by given AOI polygons.
from_boxlist(boxlist)	Make ObjectDetectionLabels from BoxList object.
from_geojson(geojson[, bbox, ioa_thresh, clip])	Convert GeoJSON to ObjectDetectionLabels object.
from_predictions(windows, predictions)	Instantiate from windows and their corresponding predictions.
get_boxes()	Return list of Boxes.
get_class_ids()
get_npboxes()
get_overlapping(labels, window[, ...])	Return subset of labels that overlap with window.
get_scores()
global_to_local(npboxes, window)	Convert from global to local coordinates.
local_to_global(npboxes, window)	Convert from local to global coordinates.
local_to_normalized(npboxes, window)	Convert from local to normalized coordinates.
make_empty()	Instantiate an empty instance of this class.
normalized_to_local(npboxes, window)	Convert from normalized to local coordinates.
prune_duplicates(labels, score_thresh, ...)	Remove duplicate boxes via non-maximum suppression.
save(uri, class_config, crs_transformer)	Save labels as a GeoJSON file.
to_boxlist()
to_dict([round_boxes])	Returns a dict version of these labels.

__add__(other: ObjectDetectionLabels) → ObjectDetectionLabels

Add labels to these labels.

Returns a concatenation of this and the other labels.

Parameters

other (ObjectDetectionLabels) –

Return type

ObjectDetectionLabels

__getitem__(window: Box) → ObjectDetectionLabels

Parameters

window (Box) –

Return type

ObjectDetectionLabels

__init__(npboxes: array, class_ids: array, scores: Optional[array] = None)

Construct a set of object detection labels.

Parameters

• npboxes (array) – float numpy array of size nx4 with cols ymin, xmin, ymax, xmax. Should be in pixel coordinates within the global frame of reference.

• class_ids (array) – int numpy array of size n with class ids

• scores (Optional[array]) – float numpy array of size n

__setitem__(window: Box, item: Dict[str, ndarray])

Parameters

• window (Box) –

• item (Dict[str, ndarray]) –

assert_equal(expected_labels: ObjectDetectionLabels)

Parameters

expected_labels (ObjectDetectionLabels) –

static concatenate(labels1: ObjectDetectionLabels, labels2: ObjectDetectionLabels) → ObjectDetectionLabels

Return concatenation of labels.

Parameters

• labels1 (ObjectDetectionLabels) – ObjectDetectionLabels

• labels2 (ObjectDetectionLabels) – ObjectDetectionLabels

Return type

ObjectDetectionLabels

filter_by_aoi(aoi_polygons: Iterable[Polygon])

Return a copy of these labels filtered by given AOI polygons.

Parameters

aoi_polygons (Iterable[Polygon]) – List of AOI polygons to filter by, in pixel coordinates.

static from_boxlist(boxlist: NpBoxList)

Make ObjectDetectionLabels from BoxList object.

Parameters

boxlist (NpBoxList) –

static from_geojson(geojson: dict, bbox: Optional[Box] = None, ioa_thresh: float = 0.8, clip: bool = True) → ObjectDetectionLabels

Convert GeoJSON to ObjectDetectionLabels object.

If bbox is provided, filter out the boxes that lie “more than a little bit” outside the bbox.

Parameters

• geojson (dict) – (dict) normalized GeoJSON (see VectorSource)

• bbox (Optional[Box]) – (Box) in pixel coords

• ioa_thresh (float) –

• clip (bool) –

Returns

ObjectDetectionLabels

Return type

ObjectDetectionLabels

classmethod from_predictions(windows: Iterable[Box], predictions: Iterable[Any]) → Labels

Instantiate from windows and their corresponding predictions.

This makes no assumptions about the type or format of the predictions. Subclasses should implement the __setitem__ method to correctly handle the predictions.

Parameters

• windows (Iterable[Box]) – Boxes in pixel coords, specifying chips in the raster.

• predictions (Iterable[Any]) – The model predictions for each chip specified by the windows.

Returns

An object of the Label subclass on which this method is called.

Return type

Labels

get_boxes() → List[Box]

Return list of Boxes.

Return type

List[Box]

get_class_ids() → ndarray

Return type

ndarray

get_npboxes() → ndarray

Return type

ndarray

static get_overlapping(labels: ObjectDetectionLabels, window: Box, ioa_thresh: float = 0.5, clip: bool = False) → ObjectDetectionLabels

Return subset of labels that overlap with window.

Parameters

• labels (ObjectDetectionLabels) – ObjectDetectionLabels

• window (Box) – Box

• ioa_thresh (float) – The minimum intersection-over-area (IOA) for a box to be considered as overlapping. For each box, IOA is defined as the area of the intersection of the box with the window over the area of the box.

• clip (bool) – If True, clip label boxes to the window.

Return type

ObjectDetectionLabels

get_scores() → ndarray

Return type

ndarray

static global_to_local(npboxes: ndarray, window: Box)

Convert from global to local coordinates.

The global coordinates are row/col within the extent of a RasterSource. The local coordinates are row/col within the window frame of reference.

Parameters

• npboxes (ndarray) –

• window (Box) –

static local_to_global(npboxes: ndarray, window: Box)

Convert from local to global coordinates.

The local coordinates are row/col within the window frame of reference. The global coordinates are row/col within the extent of a RasterSource.

Parameters

• npboxes (ndarray) –

• window (Box) –

static local_to_normalized(npboxes: ndarray, window: Box)

Convert from local to normalized coordinates.

The local coordinates are row/col within the window frame of reference. Normalized coordinates range from 0 to 1 on each (height/width) axis.

Parameters

• npboxes (ndarray) –

• window (Box) –

classmethod make_empty() → ObjectDetectionLabels

Instantiate an empty instance of this class.

Returns

An object of the Label subclass on which this method is called.

Return type

Labels

static normalized_to_local(npboxes: ndarray, window: Box)

Convert from normalized to local coordinates.

Normalized coordinates range from 0 to 1 on each (height/width) axis. The local coordinates are row/col within the window frame of reference.

Parameters

• npboxes (ndarray) –

• window (Box) –

static prune_duplicates(labels: ObjectDetectionLabels, score_thresh: float, merge_thresh: float, max_output_size: Optional[int] = None) → ObjectDetectionLabels

Remove duplicate boxes via non-maximum suppression.

Parameters

• labels (ObjectDetectionLabels) – Labels whose boxes are to be pruned.

• score_thresh (float) – Prune boxes with score less than this threshold.

• merge_thresh (float) – Prune boxes with intersection-over-union (IOU) greater than this threshold.

• max_output_size (int) – Maximum number of retained boxes. If None, this is set to len(abels). Defaults to None.

Returns

Pruned labels.

Return type

ObjectDetectionLabels

save(uri: str, class_config: ClassConfig, crs_transformer: CRSTransformer) → None

Save labels as a GeoJSON file.

Parameters

• uri (str) – URI of the output file.

• class_config (ClassConfig) – ClassConfig to map class IDs to names.

• crs_transformer (CRSTransformer) – CRSTransformer to convert from pixel-coords to map-coords before saving.

Return type

None

to_boxlist() → NpBoxList

Return type

NpBoxList

to_dict(round_boxes: bool = True) → dict

Returns a dict version of these labels.

The Dict has a Box as a key, and a tuple of (class_id, score) as the values.

Parameters

round_boxes (bool) –

Return type

dict