detectron2.structures package¶

class
detectron2.structures.
Boxes
(tensor: torch.Tensor)[source]¶ Bases:
object
This structure stores a list of boxes as a Nx4 torch.Tensor. It supports some common methods about boxes (area, clip, nonempty, etc), and also behaves like a Tensor (support indexing, to(device), .device, and iteration over all boxes)

tensor
¶ float matrix of Nx4. Each row is (x1, y1, x2, y2).
 Type

__init__
(tensor: torch.Tensor)[source]¶  Parameters
tensor (Tensor[float]) – a Nx4 matrix. Each row is (x1, y1, x2, y2).

area
() → torch.Tensor[source]¶ Computes the area of all the boxes.
 Returns
torch.Tensor – a vector with areas of each box.

clip
(box_size: Tuple[int, int]) → None[source]¶ Clip (in place) the boxes by limiting x coordinates to the range [0, width] and y coordinates to the range [0, height].
 Parameters
box_size (height, width) – The clipping box’s size.

nonempty
(threshold: float = 0.0) → torch.Tensor[source]¶ Find boxes that are nonempty. A box is considered empty, if either of its side is no larger than threshold.
 Returns
Tensor – a binary vector which represents whether each box is empty (False) or nonempty (True).

__getitem__
(item)[source]¶  Parameters
item – int, slice, or a BoolTensor
 Returns
Boxes – Create a new
Boxes
by indexing.
The following usage are allowed:
new_boxes = boxes[3]: return a Boxes which contains only one box.
new_boxes = boxes[2:10]: return a slice of boxes.
new_boxes = boxes[vector], where vector is a torch.BoolTensor with length = len(boxes). Nonzero elements in the vector will be selected.
Note that the returned Boxes might share storage with this Boxes, subject to Pytorch’s indexing semantics.

inside_box
(box_size: Tuple[int, int], boundary_threshold: int = 0) → torch.Tensor[source]¶  Parameters
box_size (height, width) – Size of the reference box.
boundary_threshold (int) – Boxes that extend beyond the reference box boundary by more than boundary_threshold are considered “outside”.
 Returns
a binary vector, indicating whether each box is inside the reference box.

get_centers
() → torch.Tensor[source]¶  Returns
The box centers in a Nx2 array of (x, y).

scale
(scale_x: float, scale_y: float) → None[source]¶ Scale the box with horizontal and vertical scaling factors

property
device
¶


class
detectron2.structures.
BoxMode
(value)[source]¶ Bases:
enum.IntEnum
Enum of different ways to represent a box.

XYXY_ABS
= 0¶ (x0, y0, x1, y1) in absolute floating points coordinates. The coordinates in range [0, width or height].

XYWH_ABS
= 1¶ (x0, y0, w, h) in absolute floating points coordinates.

XYXY_REL
= 2¶ Not yet supported! (x0, y0, x1, y1) in range [0, 1]. They are relative to the size of the image.

XYWH_REL
= 3¶ Not yet supported! (x0, y0, w, h) in range [0, 1]. They are relative to the size of the image.

XYWHA_ABS
= 4¶ (xc, yc, w, h, a) in absolute floating points coordinates. (xc, yc) is the center of the rotated box, and the angle a is in degrees ccw.

static
convert
(box: Union[List[float], Tuple[float, …], torch.Tensor, numpy.ndarray], from_mode: detectron2.structures.boxes.BoxMode, to_mode: detectron2.structures.boxes.BoxMode) → Union[List[float], Tuple[float, …], torch.Tensor, numpy.ndarray][source]¶  Parameters
box – can be a ktuple, klist or an Nxk array/tensor, where k = 4 or 5
to_mode (from_mode,) –
 Returns
The converted box of the same type.


detectron2.structures.
pairwise_iou
(boxes1: detectron2.structures.boxes.Boxes, boxes2: detectron2.structures.boxes.Boxes) → torch.Tensor[source]¶ Given two lists of boxes of size N and M, compute the IoU (intersection over union) between __all__ N x M pairs of boxes. The box order must be (xmin, ymin, xmax, ymax).
 Parameters
boxes1,boxes2 (Boxes) – two Boxes. Contains N & M boxes, respectively.
 Returns
Tensor – IoU, sized [N,M].

class
detectron2.structures.
ImageList
(tensor: torch.Tensor, image_sizes: List[Tuple[int, int]])[source]¶ Bases:
object
Structure that holds a list of images (of possibly varying sizes) as a single tensor. This works by padding the images to the same size, and storing in a field the original sizes of each image

__init__
(tensor: torch.Tensor, image_sizes: List[Tuple[int, int]])[source]¶

__getitem__
(idx) → torch.Tensor[source]¶ Access the individual image in its original size.
 Parameters
idx – int or slice
 Returns
Tensor – an image of shape (H, W) or (C_1, …, C_K, H, W) where K >= 1

property
device
¶

static
from_tensors
(tensors: Sequence[torch.Tensor], size_divisibility: int = 0, pad_value: float = 0.0) → detectron2.structures.image_list.ImageList[source]¶  Parameters
tensors – a tuple or list of torch.Tensors, each of shape (Hi, Wi) or (C_1, …, C_K, Hi, Wi) where K >= 1. The Tensors will be padded to the same shape with pad_value.
size_divisibility (int) – If size_divisibility > 0, add padding to ensure the common height and width is divisible by size_divisibility. This depends on the model and many models need a divisibility of 32.
pad_value (float) – value to pad
 Returns
an ImageList.


class
detectron2.structures.
Instances
(image_size: Tuple[int, int], **kwargs: Any)[source]¶ Bases:
object
This class represents a list of instances in an image. It stores the attributes of instances (e.g., boxes, masks, labels, scores) as “fields”. All fields must have the same
__len__
which is the number of instances.All other (nonfield) attributes of this class are considered private: they must start with ‘_’ and are not modifiable by a user.
Some basic usage:
Set/get/check a field:
instances.gt_boxes = Boxes(...) print(instances.pred_masks) # a tensor of shape (N, H, W) print('gt_masks' in instances)
len(instances)
returns the number of instancesIndexing:
instances[indices]
will apply the indexing on all the fields and returns a newInstances
. Typically,indices
is a integer vector of indices, or a binary mask of lengthnum_instances

__init__
(image_size: Tuple[int, int], **kwargs: Any)[source]¶  Parameters
image_size (height, width) – the spatial size of the image.
kwargs – fields to add to this Instances.

property
image_size
¶ Returns: tuple: height, width

set
(name: str, value: Any) → None[source]¶ Set the field named name to value. The length of value must be the number of instances, and must agree with other existing fields in this object.

get_fields
() → Dict[str, Any][source]¶  Returns
dict – a dict which maps names (str) to data of the fields
Modifying the returned dict will modify this instance.

to
(*args: Any, **kwargs: Any) → detectron2.structures.instances.Instances[source]¶  Returns
Instances – all fields are called with a to(device), if the field has this method.

__getitem__
(item: Union[int, slice, torch.BoolTensor]) → detectron2.structures.instances.Instances[source]¶  Parameters
item – an indexlike object and will be used to index all the fields.
 Returns
If item is a string, return the data in the corresponding field. Otherwise, returns an Instances where all fields are indexed by item.

class
detectron2.structures.
Keypoints
(keypoints: Union[torch.Tensor, numpy.ndarray, List[List[float]]])[source]¶ Bases:
object
Stores keypoint annotation data. GT Instances have a gt_keypoints property containing the x,y location and visibility flag of each keypoint. This tensor has shape (N, K, 3) where N is the number of instances and K is the number of keypoints per instance.
The visibility flag follows the COCO format and must be one of three integers: * v=0: not labeled (in which case x=y=0) * v=1: labeled but not visible * v=2: labeled and visible

__init__
(keypoints: Union[torch.Tensor, numpy.ndarray, List[List[float]]])[source]¶  Parameters
keypoints – A Tensor, numpy array, or list of the x, y, and visibility of each keypoint. The shape should be (N, K, 3) where N is the number of instances, and K is the number of keypoints per instance.

property
device
¶

to_heatmap
(boxes: torch.Tensor, heatmap_size: int) → torch.Tensor[source]¶ Convert keypoint annotations to a heatmap of onehot labels for training, as described in Mask RCNN.
 Parameters
boxes – Nx4 tensor, the boxes to draw the keypoints to
 Returns
 heatmaps – A tensor of shape (N, K), each element is integer spatial label
in the range [0, heatmap_size**2  1] for each keypoint in the input.
 valid:
A tensor of shape (N, K) containing whether each keypoint is in the roi or not.

__getitem__
(item: Union[int, slice, torch.BoolTensor]) → detectron2.structures.keypoints.Keypoints[source]¶ Create a new Keypoints by indexing on this Keypoints.
The following usage are allowed:
new_kpts = kpts[3]: return a Keypoints which contains only one instance.
new_kpts = kpts[2:10]: return a slice of key points.
new_kpts = kpts[vector], where vector is a torch.ByteTensor with length = len(kpts). Nonzero elements in the vector will be selected.
Note that the returned Keypoints might share storage with this Keypoints, subject to Pytorch’s indexing semantics.


detectron2.structures.
heatmaps_to_keypoints
(maps: torch.Tensor, rois: torch.Tensor) → torch.Tensor[source]¶ Extract predicted keypoint locations from heatmaps.
 Parameters
maps (Tensor) – (#ROIs, #keypoints, POOL_H, POOL_W). The predicted heatmap of logits for each ROI and each keypoint.
rois (Tensor) – (#ROIs, 4). The box of each ROI.
 Returns
Tensor of shape (#ROIs, #keypoints, 4) with the last dimension corresponding to (x, y, logit, score) for each keypoint.
When converting discrete pixel indices in an NxN image to a continuous keypoint coordinate, we maintain consistency with
Keypoints.to_heatmap()
by using the conversion from Heckbert 1990: c = d + 0.5, where d is a discrete coordinate and c is a continuous coordinate.

class
detectron2.structures.
BitMasks
(tensor: Union[torch.Tensor, numpy.ndarray])[source]¶ Bases:
object
This class stores the segmentation masks for all objects in one image, in the form of bitmaps.

tensor
¶ bool Tensor of N,H,W, representing N instances in the image.

__init__
(tensor: Union[torch.Tensor, numpy.ndarray])[source]¶  Parameters
tensor – bool Tensor of N,H,W, representing N instances in the image.

property
device
¶

__getitem__
(item: Union[int, slice, torch.BoolTensor]) → detectron2.structures.masks.BitMasks[source]¶  Returns
BitMasks – Create a new
BitMasks
by indexing.
The following usage are allowed:
new_masks = masks[3]: return a BitMasks which contains only one mask.
new_masks = masks[2:10]: return a slice of masks.
new_masks = masks[vector], where vector is a torch.BoolTensor with length = len(masks). Nonzero elements in the vector will be selected.
Note that the returned object might share storage with this object, subject to Pytorch’s indexing semantics.

nonempty
() → torch.Tensor[source]¶ Find masks that are nonempty.
 Returns
Tensor –
 a BoolTensor which represents
whether each mask is empty (False) or nonempty (True).

static
from_polygon_masks
(polygon_masks: Union[PolygonMasks, List[List[numpy.ndarray]]], height: int, width: int) → detectron2.structures.masks.BitMasks[source]¶  Parameters
polygon_masks (list[list[ndarray]] or PolygonMasks) –
width (height,) –

crop_and_resize
(boxes: torch.Tensor, mask_size: int) → torch.Tensor[source]¶ Crop each bitmask by the given box, and resize results to (mask_size, mask_size). This can be used to prepare training targets for Mask RCNN. It has less reconstruction error compared to rasterization with polygons. However we observe no difference in accuracy, but BitMasks requires more memory to store all the masks.
 Parameters
boxes (Tensor) – Nx4 tensor storing the boxes for each mask
mask_size (int) – the size of the rasterized mask.
 Returns
Tensor – A bool tensor of shape (N, mask_size, mask_size), where N is the number of predicted boxes for this image.


class
detectron2.structures.
PolygonMasks
(polygons: List[List[Union[torch.Tensor, numpy.ndarray]]])[source]¶ Bases:
object
This class stores the segmentation masks for all objects in one image, in the form of polygons.

polygons
¶ list[list[ndarray]]. Each ndarray is a float64 vector representing a polygon.

__init__
(polygons: List[List[Union[torch.Tensor, numpy.ndarray]]])[source]¶  Parameters
polygons (list[list[np.ndarray]]) – The first level of the list correspond to individual instances, the second level to all the polygons that compose the instance, and the third level to the polygon coordinates. The third level array should have the format of [x0, y0, x1, y1, …, xn, yn] (n >= 3).

property
device
¶

get_bounding_boxes
() → detectron2.structures.boxes.Boxes[source]¶  Returns
Boxes – tight bounding boxes around polygon masks.

nonempty
() → torch.Tensor[source]¶ Find masks that are nonempty.
 Returns
Tensor – a BoolTensor which represents whether each mask is empty (False) or not (True).

__getitem__
(item: Union[int, slice, List[int], torch.BoolTensor]) → detectron2.structures.masks.PolygonMasks[source]¶ Support indexing over the instances and return a PolygonMasks object. item can be:
An integer. It will return an object with only one instance.
A slice. It will return an object with the selected instances.
A list[int]. It will return an object with the selected instances, correpsonding to the indices in the list.
A vector mask of type BoolTensor, whose length is num_instances. It will return an object with the instances whose mask is nonzero.

__iter__
() → Iterator[List[numpy.ndarray]][source]¶  Yields
list[ndarray] – the polygons for one instance. Each Tensor is a float64 vector representing a polygon.

crop_and_resize
(boxes: torch.Tensor, mask_size: int) → torch.Tensor[source]¶ Crop each mask by the given box, and resize results to (mask_size, mask_size). This can be used to prepare training targets for Mask RCNN.
 Parameters
boxes (Tensor) – Nx4 tensor storing the boxes for each mask
mask_size (int) – the size of the rasterized mask.
 Returns
Tensor – A bool tensor of shape (N, mask_size, mask_size), where N is the number of predicted boxes for this image.

area
()[source]¶ Computes area of the mask. Only works with Polygons, using the shoelace formula: https://stackoverflow.com/questions/24467972/calculateareaofpolygongivenxycoordinates
 Returns
Tensor – a vector, area for each instance

static
cat
(polymasks_list: List[PolygonMasks]) → detectron2.structures.masks.PolygonMasks[source]¶ Concatenates a list of PolygonMasks into a single PolygonMasks
 Parameters
polymasks_list (list[PolygonMasks]) –
 Returns
PolygonMasks – the concatenated PolygonMasks


detectron2.structures.
rasterize_polygons_within_box
(polygons: List[numpy.ndarray], box: numpy.ndarray, mask_size: int) → torch.Tensor[source]¶ Rasterize the polygons into a mask image and crop the mask content in the given box. The cropped mask is resized to (mask_size, mask_size).
This function is used when generating training targets for mask head in Mask RCNN. Given original groundtruth masks for an image, new groundtruth mask training targets in the size of mask_size x mask_size must be provided for each predicted box. This function will be called to produce such targets.

detectron2.structures.
polygons_to_bitmask
(polygons: List[numpy.ndarray], height: int, width: int) → numpy.ndarray[source]¶  Parameters
polygons (list[ndarray]) – each array has shape (Nx2,)
width (height,) –
 Returns
ndarray – a bool mask of shape (height, width)

class
detectron2.structures.
RotatedBoxes
(tensor: torch.Tensor)[source]¶ Bases:
detectron2.structures.boxes.Boxes
This structure stores a list of rotated boxes as a Nx5 torch.Tensor. It supports some common methods about boxes (area, clip, nonempty, etc), and also behaves like a Tensor (support indexing, to(device), .device, and iteration over all boxes)

__init__
(tensor: torch.Tensor)[source]¶  Parameters
tensor (Tensor[float]) – a Nx5 matrix. Each row is (x_center, y_center, width, height, angle), in which angle is represented in degrees. While there’s no strict range restriction for it, the recommended principal range is between [180, 180) degrees.
Assume we have a horizontal box B = (x_center, y_center, width, height), where width is along the xaxis and height is along the yaxis. The rotated box B_rot (x_center, y_center, width, height, angle) can be seen as:
When angle == 0: B_rot == B
When angle > 0: B_rot is obtained by rotating B w.r.t its center by \(angle\) degrees CCW;
When angle < 0: B_rot is obtained by rotating B w.r.t its center by \(angle\) degrees CW.
Mathematically, since the righthanded coordinate system for image space is (y, x), where y is top>down and x is left>right, the 4 vertices of the rotated rectangle \((yr_i, xr_i)\) (i = 1, 2, 3, 4) can be obtained from the vertices of the horizontal rectangle (y_i, x_i) (i = 1, 2, 3, 4) in the following way (\(\theta = angle*\pi/180\) is the angle in radians, (y_c, x_c) is the center of the rectangle):
\[ \begin{align}\begin{aligned}yr_i = \cos(\theta) (y_i  y_c)  \sin(\theta) (x_i  x_c) + y_c,\\xr_i = \sin(\theta) (y_i  y_c) + \cos(\theta) (x_i  x_c) + x_c,\end{aligned}\end{align} \]which is the standard rigidbody rotation transformation.
Intuitively, the angle is (1) the rotation angle from yaxis in image space to the height vector (top>down in the box’s local coordinate system) of the box in CCW, and (2) the rotation angle from xaxis in image space to the width vector (left>right in the box’s local coordinate system) of the box in CCW.
More intuitively, consider the following horizontal box ABCD represented in (x1, y1, x2, y2): (3, 2, 7, 4), covering the [3, 7] x [2, 4] region of the continuous coordinate system which looks like this:
O> x   AB     DC  v y
Note that each capital letter represents one 0dimensional geometric point instead of a ‘square pixel’ here.
In the example above, using (x, y) to represent a point we have:
\[O = (0, 0), A = (3, 2), B = (7, 2), C = (7, 4), D = (3, 4)\]We name vector AB = vector DC as the width vector in box’s local coordinate system, and vector AD = vector BC as the height vector in box’s local coordinate system. Initially, when angle = 0 degree, they’re aligned with the positive directions of xaxis and yaxis in the image space, respectively.
For better illustration, we denote the center of the box as E,
O> x   AB   E   DC  v y
where the center E = ((3+7)/2, (2+4)/2) = (5, 3).
Also,
\[width = AB = CD = 7  3 = 4, height = AD = BC = 4  2 = 2.\]Therefore, the corresponding representation for the same shape in rotated box in (x_center, y_center, width, height, angle) format is:
(5, 3, 4, 2, 0),
Now, let’s consider (5, 3, 4, 2, 90), which is rotated by 90 degrees CCW (counterclockwise) by definition. It looks like this:
O> x  BC     E     AD v y
The center E is still located at the same point (5, 3), while the vertices ABCD are rotated by 90 degrees CCW with regard to E: A = (4, 5), B = (4, 1), C = (6, 1), D = (6, 5)
Here, 90 degrees can be seen as the CCW angle to rotate from yaxis to vector AD or vector BC (the top>down height vector in box’s local coordinate system), or the CCW angle to rotate from xaxis to vector AB or vector DC (the left>right width vector in box’s local coordinate system).
\[width = AB = CD = 5  1 = 4, height = AD = BC = 6  4 = 2.\]Next, how about (5, 3, 4, 2, 90), which is rotated by 90 degrees CW (clockwise) by definition? It looks like this:
O> x  DA     E     CB v y
The center E is still located at the same point (5, 3), while the vertices ABCD are rotated by 90 degrees CW with regard to E: A = (6, 1), B = (6, 5), C = (4, 5), D = (4, 1)
\[width = AB = CD = 5  1 = 4, height = AD = BC = 6  4 = 2.\]This covers exactly the same region as (5, 3, 4, 2, 90) does, and their IoU will be 1. However, these two will generate different RoI Pooling results and should not be treated as an identical box.
On the other hand, it’s easy to see that (X, Y, W, H, A) is identical to (X, Y, W, H, A+360N), for any integer N. For example (5, 3, 4, 2, 270) would be identical to (5, 3, 4, 2, 90), because rotating the shape 270 degrees CCW is equivalent to rotating the same shape 90 degrees CW.
We could rotate further to get (5, 3, 4, 2, 180), or (5, 3, 4, 2, 180):
O> x   CD   E   BA  v y
\[ \begin{align}\begin{aligned}A = (7, 4), B = (3, 4), C = (3, 2), D = (7, 2),\\width = AB = CD = 7  3 = 4, height = AD = BC = 4  2 = 2.\end{aligned}\end{align} \]Finally, this is a very inaccurate (heavily quantized) illustration of how (5, 3, 4, 2, 60) looks like in case anyone wonders:
O> x  B  / C  /E /  A /  `D v y
It’s still a rectangle with center of (5, 3), width of 4 and height of 2, but its angle (and thus orientation) is somewhere between (5, 3, 4, 2, 0) and (5, 3, 4, 2, 90).

clone
() → detectron2.structures.rotated_boxes.RotatedBoxes[source]¶ Clone the RotatedBoxes.
 Returns
RotatedBoxes

area
() → torch.Tensor[source]¶ Computes the area of all the boxes.
 Returns
torch.Tensor – a vector with areas of each box.

clip
(box_size: Tuple[int, int], clip_angle_threshold: float = 1.0) → None[source]¶ Clip (in place) the boxes by limiting x coordinates to the range [0, width] and y coordinates to the range [0, height].
For RRPN: Only clip boxes that are almost horizontal with a tolerance of clip_angle_threshold to maintain backward compatibility.
Rotated boxes beyond this threshold are not clipped for two reasons:
There are potentially multiple ways to clip a rotated box to make it fit within the image.
It’s tricky to make the entire rectangular box fit within the image and still be able to not leave out pixels of interest.
Therefore we rely on ops like RoIAlignRotated to safely handle this.
 Parameters
box_size (height, width) – The clipping box’s size.
clip_angle_threshold – Iff. abs(normalized(angle)) <= clip_angle_threshold (in degrees), we do the clipping as horizontal boxes.

nonempty
(threshold: float = 0.0) → torch.Tensor[source]¶ Find boxes that are nonempty. A box is considered empty, if either of its side is no larger than threshold.
 Returns
Tensor – a binary vector which represents whether each box is empty (False) or nonempty (True).

__getitem__
(item: Union[int, slice, torch.BoolTensor]) → detectron2.structures.rotated_boxes.RotatedBoxes[source]¶  Returns
RotatedBoxes – Create a new
RotatedBoxes
by indexing.
The following usage are allowed:
new_boxes = boxes[3]: return a RotatedBoxes which contains only one box.
new_boxes = boxes[2:10]: return a slice of boxes.
new_boxes = boxes[vector], where vector is a torch.ByteTensor with length = len(boxes). Nonzero elements in the vector will be selected.
Note that the returned RotatedBoxes might share storage with this RotatedBoxes, subject to Pytorch’s indexing semantics.

inside_box
(box_size: Tuple[int, int], boundary_threshold: int = 0) → torch.Tensor[source]¶  Parameters
box_size (height, width) – Size of the reference box covering [0, width] x [0, height]
boundary_threshold (int) – Boxes that extend beyond the reference box boundary by more than boundary_threshold are considered “outside”.
For RRPN, it might not be necessary to call this function since it’s common for rotated box to extend to outside of the image boundaries (the clip function only clips the nearhorizontal boxes)
 Returns
a binary vector, indicating whether each box is inside the reference box.

get_centers
() → torch.Tensor[source]¶  Returns
The box centers in a Nx2 array of (x, y).

scale
(scale_x: float, scale_y: float) → None[source]¶ Scale the rotated box with horizontal and vertical scaling factors Note: when scale_factor_x != scale_factor_y, the rotated box does not preserve the rectangular shape when the angle is not a multiple of 90 degrees under resize transformation. Instead, the shape is a parallelogram (that has skew) Here we make an approximation by fitting a rotated rectangle to the parallelogram.

property
device
¶

__iter__
() → Iterator[torch.Tensor][source]¶ Yield a box as a Tensor of shape (5,) at a time.


detectron2.structures.
pairwise_iou_rotated
(boxes1: detectron2.structures.rotated_boxes.RotatedBoxes, boxes2: detectron2.structures.rotated_boxes.RotatedBoxes) → None¶ Given two lists of rotated boxes of size N and M, compute the IoU (intersection over union) between __all__ N x M pairs of boxes. The box order must be (x_center, y_center, width, height, angle).
 Parameters
boxes2 (boxes1,) – two RotatedBoxes. Contains N & M rotated boxes, respectively.
 Returns
Tensor – IoU, sized [N,M].