from typing import Union, Tuple, List
import numpy as np
import torch
from ..utils import EPS
from ..utils.decorators import np2torch, shape_check
[docs]
@np2torch(dtype=torch.int)
@shape_check(2)
def miou(img1: np.ndarray | torch.Tensor, img2: np.ndarray | torch.Tensor) -> torch.Tensor:
"""
Compute the mean Intersection over Union (mIoU) for each value in the given images.
This function calculates the mIoU for 1-dimensional arrays or tensors.
Future updates will extend support for higher-dimensional arrays or tensors.
Parameters
----------
img1 : Union[np.ndarray, torch.Tensor]
The first input image, either as a numpy array or a torch tensor. Shape should be [-1, dim].
img2 : Union[np.ndarray, torch.Tensor]
The second input image, either as a numpy array or a torch tensor. Shape should be [-1, dim].
Returns
-------
torch.Tensor
The mIoU values for each class, as a torch tensor. Shape is [-1].
Examples
--------
>>> img1 = np.array([0, 1, 1, 2, 2])
>>> img2 = np.array([0, 1, 1, 2, 1])
>>> miou(img1, img2)
tensor([1.0000, 0.6667, 0.0000])
>>> img1 = torch.tensor([0, 1, 1, 2, 2])
>>> img2 = torch.tensor([0, 1, 1, 2, 1])
>>> miou(img1, img2)
tensor([1.0000, 0.6667, 0.0000])
"""
# TODO: this function works for 1-dimensional arrays or tensors, 2 or greater dimensional mode will be update.
if len(img1.shape) == 1:
img1 = img1.clone().reshape(1, -1)
img2 = img2.clone().reshape(1, -1)
n = len(img1)
maximum = int(max(img1.max(), img2.max()))
iou_ = torch.zeros(n)
for m in range(maximum):
intersection = ((img1 == m).int() & (img2 == m).int()).float().sum(-1)
union = ((img1 == m).int() | (img2 == m).int()).float().sum(-1)
iou_ += intersection / (union + EPS)
return iou_ / maximum
[docs]
@np2torch(dtype=torch.int)
@shape_check("img1", "img2")
def giou(
img1: np.ndarray | torch.Tensor,
img2: np.ndarray | torch.Tensor,
iou: bool = False,
) -> Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]:
"""
Compute the Generalized Intersection over Union (GIoU) and optionally the Intersection over Union (IoU) for given images.
This function calculates the GIoU and optionally the IoU for bounding boxes in object detection tasks.
Parameters
----------
img1 : Union[np.ndarray, torch.Tensor]
The first input image, either as a numpy array or a torch tensor. Shape should be [-1, 2], representing on-off pairs.
img2 : Union[np.ndarray, torch.Tensor]
The second input image, either as a numpy array or a torch tensor. Shape should be [-1, 2], representing on-off pairs.
iou : bool, optional
If True, return both IoU and GIoU. Default is False.
Returns
-------
Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]
The GIoU values, and optionally the IoU values, as torch tensors. Shape is [-1].
Examples
--------
>>> img1 = np.array([[1, 4], [2, 5]])
>>> img2 = np.array([[1, 3], [2, 6]])
>>> giou(img1, img2)
tensor([-0.6667, -0.5000])
>>> giou(img1, img2, iou=True)
(tensor([-0.6667, -0.5000]), tensor([0.5000, 0.3333]))
>>> img1 = torch.tensor([[1, 4], [2, 5]])
>>> img2 = torch.tensor([[1, 3], [2, 6]])
>>> giou(img1, img2)
tensor([-0.6667, -0.5000])
>>> giou(img1, img2, iou=True)
(tensor([-0.6667, -0.5000]), tensor([0.5000, 0.3333]))
"""
if len(img1.shape) == 1:
img1 = img1.clone().unsqueeze(0)
if len(img2.shape) == 1:
img2 = img2.clone().unsqueeze(0)
true_on, true_off = img1[:, 0], img1[:, 1]
pred_on, pred_off = img2[:, 0], img2[:, 1]
C_on = torch.min(true_on, pred_on)
C_off = torch.max(true_off, pred_off)
C_area = C_off - C_on
relative_area = C_area - (true_off - true_on)
union = C_area # they are same in 1-dimension
intersection = torch.min(abs(true_on - pred_off), abs(true_off - pred_on))
IoU = intersection / (union + EPS)
if iou:
return IoU - abs(relative_area / (C_area + EPS)), IoU
else:
return IoU - abs(relative_area / (C_area + EPS))