Distance IoU

docs/source/ops.rst

@@ -24,6 +24,8 @@ Operators
     drop_block3d
     generalized_box_iou
     generalized_box_iou_loss
+    distance_box_iou
+    distance_box_iou_loss
     masks_to_boxes
     nms
     ps_roi_align

test/test_ops.py

@@ -1258,6 +1258,78 @@ def test_giou_jit(self) -> None:
         self._run_jit_test([[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]])
 
 
+class TestDistanceBoxIoU(BoxTestBase):
+    def _target_fn(self):
+        return (True, ops.distance_box_iou)
+
+    def _generate_int_input():
+        return [[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]]
+
+    def _generate_int_expected():
+        return [[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0]]
+
+    def _generate_float_input():
+        return [
+            [285.3538, 185.5758, 1193.5110, 851.4551],
+            [285.1472, 188.7374, 1192.4984, 851.0669],
+            [279.2440, 197.9812, 1189.4746, 849.2019],
+        ]
+
+    def _generate_float_expected():
+        return [[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]]
+
+    @pytest.mark.parametrize(
+        "test_input, dtypes, tolerance, expected",
+        [
+            pytest.param(
+                _generate_int_input(), [torch.int16, torch.int32, torch.int64], 1e-4, _generate_int_expected()
+            ),
+            pytest.param(_generate_float_input(), [torch.float16], 0.002, _generate_float_expected()),
+            pytest.param(_generate_float_input(), [torch.float32, torch.float64], 0.001, _generate_float_expected()),
+        ],
+    )
+    def test_distance_iou(self, test_input, dtypes, tolerance, expected):
+        self._run_test(test_input, dtypes, tolerance, expected)
+
+    def test_distance_iou_jit(self):
+        self._run_jit_test([[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]])
+
+
+@pytest.mark.parametrize("dtype", [torch.float32, torch.half])
+@pytest.mark.parametrize("device", cpu_and_gpu())
+def test_distance_iou_loss(dtype, device):
+    box1 = torch.tensor([-1, -1, 1, 1], dtype=dtype, device=device)
+    box2 = torch.tensor([0, 0, 1, 1], dtype=dtype, device=device)
+    box3 = torch.tensor([0, 1, 1, 2], dtype=dtype, device=device)
+    box4 = torch.tensor([1, 1, 2, 2], dtype=dtype, device=device)
+
+    box1s = torch.stack(
+        [box2, box2],
+        dim=0,
+    )
+    box2s = torch.stack(
+        [box3, box4],
+        dim=0,
+    )
+
+    def assert_distance_iou_loss(box1, box2, expected_output, reduction="none"):
+        output = ops.distance_box_iou_loss(box1, box2, reduction=reduction)
+        expected_output = torch.tensor(expected_output, dtype=dtype, device=device)
+        tol = 1e-5 if dtype != torch.half else 1e-3
+        torch.testing.assert_close(output, expected_output, rtol=tol, atol=tol)
+
+    assert_distance_iou_loss(box1, box1, 0.0)
+
+    assert_distance_iou_loss(box1, box2, 0.8750)
+
+    assert_distance_iou_loss(box1, box3, 1.1923)
+
+    assert_distance_iou_loss(box1, box4, 1.2778)
+
+    assert_distance_iou_loss(box1s, box2s, 1.9000, reduction="mean")
+    assert_distance_iou_loss(box1s, box2s, 3.8000, reduction="sum")
+
+
 class TestMasksToBoxes:
     def test_masks_box(self):
         def masks_box_check(masks, expected, tolerance=1e-4):

torchvision/ops/__init__.py

@@ -7,10 +7,12 @@
     box_area,
     box_iou,
     generalized_box_iou,
+    distance_box_iou,
     masks_to_boxes,
 )
 from .boxes import box_convert
 from .deform_conv import deform_conv2d, DeformConv2d
+from .diou_loss import distance_box_iou_loss
 from .drop_block import drop_block2d, DropBlock2d, drop_block3d, DropBlock3d
 from .feature_pyramid_network import FeaturePyramidNetwork
 from .focal_loss import sigmoid_focal_loss
@@ -38,6 +40,7 @@
     "box_area",
     "box_iou",
     "generalized_box_iou",
+    "distance_box_iou",
     "roi_align",
     "RoIAlign",
     "roi_pool",
@@ -56,6 +59,7 @@
     "Conv3dNormActivation",
     "SqueezeExcitation",
     "generalized_box_iou_loss",
+    "distance_box_iou_loss",
     "drop_block2d",
     "DropBlock2d",
     "drop_block3d",

torchvision/ops/boxes.py

@@ -311,6 +311,50 @@ def generalized_box_iou(boxes1: Tensor, boxes2: Tensor) -> Tensor:
     return iou - (areai - union) / areai
 
 
+def distance_box_iou(boxes1: Tensor, boxes2: Tensor, eps: float = 1e-7) -> Tensor:
+    """
+    Return distance intersection-over-union (Jaccard index) between two sets of boxes.
+
+    Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
+    ``0 <= x1 < x2`` and ``0 <= y1 < y2``.
+
+    Args:
+        boxes1 (Tensor[N, 4]): first set of boxes
+        boxes2 (Tensor[M, 4]): second set of boxes
+        eps (float, optional): small number to prevent division by zero. Default: 1e-7
+
+    Returns:
+        Tensor[N, M]: the NxM matrix containing the pairwise distance IoU values
+        for every element in boxes1 and boxes2
+    """
+    if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+        _log_api_usage_once(distance_box_iou)
+
+    boxes1 = _upcast(boxes1)
+    boxes2 = _upcast(boxes2)
+
+    inter, union = _box_inter_union(boxes1, boxes2)
+    iou = inter / union
+
+    lti = torch.min(boxes1[:, None, :2], boxes2[:, None, :2])
+    rbi = torch.max(boxes1[:, None, 2:], boxes2[:, None, 2:])
+
+    whi = _upcast(rbi - lti).clamp(min=0)  # [N,M,2]
+    diagonal_distance_squared = (whi[:, :, 0] ** 2) + (whi[:, :, 1] ** 2) + eps
+
+    # centers of boxes
+    x_p = boxes1[:, None, :2].sum() / 2
+    y_p = boxes1[:, None, 2:].sum() / 2
+    x_g = boxes2[:, None, :2].sum() / 2
+    y_g = boxes2[:, None, 2:].sum() / 2
+    # The distance between boxes' centers squared.
+    centers_distance_squared = (_upcast(x_p - x_g) ** 2) + (_upcast(y_p - y_g) ** 2)
+
+    # The distance IoU is the IoU penalized by a normalized
+    # distance between boxes' centers squared.
+    return iou - (centers_distance_squared / diagonal_distance_squared)
+
+
 def masks_to_boxes(masks: torch.Tensor) -> torch.Tensor:
     """
     Compute the bounding boxes around the provided masks.

torchvision/ops/diou_loss.py

@@ -0,0 +1,51 @@
+import torch
+
+from .boxes import distance_box_iou
+
+
+def distance_box_iou_loss(
+    boxes1: torch.Tensor,
+    boxes2: torch.Tensor,
+    reduction: str = "none",
+    eps: float = 1e-7,
+) -> torch.Tensor:
+    """
+    Gradient-friendly IoU loss with an additional penalty that is non-zero when the
+    distance between boxes' centers isn't zero. Indeed, for two exactly overlapping
+    boxes, the distance IoU is the same as the IoU loss.
+    This loss is symmetric, so the boxes1 and boxes2 arguments are interchangeable.
+
+    Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
+    ``0 <= x1 < x2`` and ``0 <= y1 < y2``, and The two boxes should have the
+    same dimensions.
+
+    Args:
+        boxes1 (Tensor[N, 4] or Tensor[4]): first set of boxes
+        boxes2 (Tensor[N, 4] or Tensor[4]): second set of boxes
+        reduction (string, optional): Specifies the reduction to apply to the output:
+            ``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: No reduction will be
+            applied to the output. ``'mean'``: The output will be averaged.
+            ``'sum'``: The output will be summed. Default: ``'none'``
+        eps (float, optional): small number to prevent division by zero. Default: 1e-7
+
+    Returns:
+        Tensor[]: Loss tensor with the reduction option applied.
+
+    Reference:
+        Zhaohui Zheng et. al: Distance Intersection over Union Loss:
+        https://arxiv.org/abs/1911.08287
+    """
+    if boxes1.dim() == 1 and boxes2.dim() == 1:
+        batch_boxes1 = boxes1.unsqueeze(0)
+        batch_boxes2 = boxes2.unsqueeze(0)
+        diou = distance_box_iou(batch_boxes1, batch_boxes2, eps)[0, 0]
+    else:
+        diou = distance_box_iou(boxes1, boxes2, eps)[0]
+    loss = 1 - diou
+    if reduction == "mean":
+        loss = loss.mean() if loss.numel() > 0 else 0.0 * loss.sum()
+    elif reduction == "sum":
+        loss = loss.sum()
+    # Cast the loss to the same dtype as the input boxes
+    loss = loss.to(boxes1.dtype)
+    return loss

torchvision/ops/giou_loss.py

@@ -36,7 +36,7 @@ def generalized_box_iou_loss(
             ``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: No reduction will be
             applied to the output. ``'mean'``: The output will be averaged.
             ``'sum'``: The output will be summed. Default: ``'none'``
-        eps (float, optional): small number to prevent division by zero. Default: 1e-7
+        eps (float): small number to prevent division by zero. Default: 1e-7
 
     Reference:
         Hamid Rezatofighi et. al: Generalized Intersection over Union: