# Copyright (C) 2019-2022, François-Guillaume Fernandez.
# This program is licensed under the Apache License 2.0.
# See LICENSE or go to <https://www.apache.org/licenses/LICENSE-2.0> for full license details.
import math
from typing import Tuple
import torch
from torch import Tensor
from torchvision.ops.boxes import box_area, box_iou
__all__ = ["box_giou", "diou_loss", "ciou_loss"]
def _box_iou(boxes1: Tensor, boxes2: Tensor) -> Tuple[Tensor, Tensor]:
# from https://github.com/facebookresearch/detr/blob/master/util/box_ops.py
area1 = box_area(boxes1)
area2 = box_area(boxes2)
lt = torch.max(boxes1[:, None, :2], boxes2[:, :2]) # [N,M,2]
rb = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) # [N,M,2]
wh = (rb - lt).clamp(min=0) # [N,M,2]
inter = wh[:, :, 0] * wh[:, :, 1] # [N,M]
union = area1[:, None] + area2 - inter
iou = inter / union
return iou, union
[docs]
def box_giou(boxes1: Tensor, boxes2: Tensor) -> Tensor:
r"""Computes the Generalized-IoU as described in `"Generalized Intersection over Union: A Metric and A Loss
for Bounding Box Regression" <https://arxiv.org/pdf/1902.09630.pdf>`_. This implementation was adapted
from https://github.com/facebookresearch/detr/blob/master/util/box_ops.py
The generalized IoU is defined as follows:
.. math::
GIoU = IoU - \frac{|C - A \cup B|}{|C|}
where :math:`IoU` is the Intersection over Union,
:math:`A \cup B` is the area of the boxes' union,
and :math:`C` is the area of the smallest enclosing box covering the two boxes.
Args:
boxes1 (torch.Tensor[M, 4]): bounding boxes
boxes2 (torch.Tensor[N, 4]): bounding boxes
Returns:
torch.Tensor[M, N]: Generalized-IoU
"""
# degenerate boxes gives inf / nan results
# so do an early check
if torch.any(boxes1[:, 2:] < boxes1[:, :2]) or torch.any(boxes2[:, 2:] < boxes2[:, :2]):
raise AssertionError("Incorrect coordinate format")
iou, union = _box_iou(boxes1, boxes2)
lt = torch.min(boxes1[:, None, :2], boxes2[:, :2])
rb = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])
wh = (rb - lt).clamp(min=0) # [N,M,2]
area = wh[:, :, 0] * wh[:, :, 1]
return iou - (area - union) / area
def iou_penalty(boxes1: Tensor, boxes2: Tensor) -> Tensor:
"""Implements the penalty term for the Distance-IoU loss
Args:
boxes1 (torch.Tensor[M, 4]): bounding boxes
boxes2 (torch.Tensor[N, 4]): bounding boxes
Returns:
torch.Tensor[M, N]: penalty term
"""
# Diagonal length of the smallest enclosing box
c2 = torch.zeros((boxes1.shape[0], boxes2.shape[0], 2), device=boxes1.device)
# Assign bottom right coords
c2[..., 0] = torch.max(boxes1[:, 2].unsqueeze(-1), boxes2[:, 2].unsqueeze(-2))
c2[..., 1] = torch.max(boxes1[:, 3].unsqueeze(-1), boxes2[:, 3].unsqueeze(-2))
# Subtract top left coords
c2[..., 0].sub_(torch.min(boxes1[:, 0].unsqueeze(-1), boxes2[:, 0].unsqueeze(-2)))
c2[..., 1].sub_(torch.min(boxes1[:, 1].unsqueeze(-1), boxes2[:, 1].unsqueeze(-2)))
c2.pow_(2)
c2 = c2.sum(dim=-1)
# L2 - distance between box centers
center_dist2 = torch.zeros((boxes1.shape[0], boxes2.shape[0], 2), device=boxes1.device)
# Centers of boxes1
center_dist2[..., 0] = boxes1[:, [0, 2]].sum(dim=1).unsqueeze(1)
center_dist2[..., 1] = boxes1[:, [1, 3]].sum(dim=1).unsqueeze(1)
# Centers of boxes2
center_dist2[..., 0].sub_(boxes2[:, [0, 2]].sum(dim=1).unsqueeze(0))
center_dist2[..., 1].sub_(boxes2[:, [1, 3]].sum(dim=1).unsqueeze(0))
center_dist2.pow_(2)
center_dist2 = center_dist2.sum(dim=-1) / 4
return center_dist2 / c2
[docs]
def diou_loss(boxes1: Tensor, boxes2: Tensor) -> Tensor:
r"""Computes the Distance-IoU loss as described in `"Distance-IoU Loss: Faster and Better Learning for
Bounding Box Regression" <https://arxiv.org/pdf/1911.08287.pdf>`_.
The loss is defined as follows:
.. math::
\mathcal{L}_{DIoU} = 1 - IoU + \frac{\rho^2(b, b^{GT})}{c^2}
where :math:`IoU` is the Intersection over Union,
:math:`b` and :math:`b^{GT}` are the centers of the box and the ground truth box respectively,
:math:`c` c is the diagonal length of the smallest enclosing box covering the two boxes,
and :math:`\rho(.)` is the Euclidean distance.
.. image:: https://github.com/frgfm/Holocron/releases/download/v0.1.3/diou_loss.png
:align: center
Args:
boxes1 (torch.Tensor[M, 4]): bounding boxes
boxes2 (torch.Tensor[N, 4]): bounding boxes
Returns:
torch.Tensor[M, N]: Distance-IoU loss
"""
return 1 - box_iou(boxes1, boxes2) + iou_penalty(boxes1, boxes2)
def aspect_ratio(boxes: Tensor) -> Tensor:
"""Computes the aspect ratio of boxes
Args:
boxes (torch.Tensor[N, 4]): bounding boxes
Returns:
torch.Tensor[N]: aspect ratio
"""
return torch.atan((boxes[:, 2] - boxes[:, 0]) / (boxes[:, 3] - boxes[:, 1]))
def aspect_ratio_consistency(boxes1: Tensor, boxes2: Tensor) -> Tensor:
"""Computes the aspect ratio consistency from the complete IoU loss
Args:
boxes1 (torch.Tensor[M, 4]): bounding boxes
boxes2 (torch.Tensor[N, 4]): bounding boxes
Returns:
torch.Tensor[M, N]: aspect ratio consistency
"""
v = aspect_ratio(boxes1).unsqueeze(-1) - aspect_ratio(boxes2).unsqueeze(-2)
v.pow_(2)
v.mul_(4 / math.pi**2)
return v
[docs]
def ciou_loss(boxes1: Tensor, boxes2: Tensor) -> Tensor:
r"""Computes the Complete IoU loss as described in `"Distance-IoU Loss: Faster and Better Learning for
Bounding Box Regression" <https://arxiv.org/pdf/1911.08287.pdf>`_.
The loss is defined as follows:
.. math::
\mathcal{L}_{CIoU} = 1 - IoU + \frac{\rho^2(b, b^{GT})}{c^2} + \alpha v
where :math:`IoU` is the Intersection over Union,
:math:`b` and :math:`b^{GT}` are the centers of the box and the ground truth box respectively,
:math:`c` c is the diagonal length of the smallest enclosing box covering the two boxes,
:math:`\rho(.)` is the Euclidean distance,
:math:`\alpha` is a positive trade-off parameter,
and :math:`v` is the aspect ratio consistency.
More specifically:
.. math::
v = \frac{4}{\pi^2} \Big(\arctan{\frac{w^{GT}}{h^{GT}}} - \arctan{\frac{w}{h}}\Big)^2
and
.. math::
\alpha = \frac{v}{(1 - IoU) + v}
Args:
boxes1 (torch.Tensor[M, 4]): bounding boxes
boxes2 (torch.Tensor[N, 4]): bounding boxes
Returns:
torch.Tensor[M, N]: Complete IoU loss
Example:
>>> import torch
>>> from holocron.ops.boxes import box_ciou
>>> boxes1 = torch.tensor([[0, 0, 100, 100], [100, 100, 200, 200]], dtype=torch.float32)
>>> boxes2 = torch.tensor([[50, 50, 150, 150]], dtype=torch.float32)
>>> box_ciou(boxes1, boxes2)
"""
iou = box_iou(boxes1, boxes2)
v = aspect_ratio_consistency(boxes1, boxes2)
ciou_loss = 1 - iou + iou_penalty(boxes1, boxes2)
# Check
_filter = (v != 0) & (iou != 0)
ciou_loss[_filter].addcdiv_(v[_filter], 1 - iou[_filter] + v[_filter])
return ciou_loss