Skip to content

Fix d/c IoU for different batch sizes #6338

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
datumbox merged 13 commits into from
Aug 2, 2022
Merged

Fix d/c IoU for different batch sizes #6338

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
63c5551
Fix bug in calculating cIoU for unequal sizes
oke-aditya Jul 31, 2022
73b907d
Merge branch 'main' of github.com:pytorch/vision into fix_g/c_iou
oke-aditya Jul 31, 2022
d4bc35e
Remove comment
oke-aditya Jul 31, 2022
5777bfb
what the epsilon?
oke-aditya Jul 31, 2022
c5c39b3
Fixing DIoU
datumbox Aug 1, 2022
7bef3d0
Optimization by Francisco.
datumbox Aug 1, 2022
2936153
Fix the expected values on CompleteBoxIoU
datumbox Aug 1, 2022
13d1961
Apply suggestions from code review
datumbox Aug 1, 2022
97298f6
Merge branch 'main' into fix_g/c_iou
datumbox Aug 1, 2022
c69ce0c
Adding cartesian product test.
datumbox Aug 2, 2022
26aba8d
Merge remote-tracking branch 'oke-aditya/fix_g/c_iou' into oke_fix_g/…
datumbox Aug 2, 2022
aa1f992
remove static
datumbox Aug 2, 2022
1ba1c69
Merge branch 'main' into fix_g/c_iou
datumbox Aug 2, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
95 changes: 54 additions & 41 deletions test/test_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1111,14 +1111,6 @@ def test_bbox_convert_jit(self):
torch.testing.assert_close(scripted_cxcywh, box_cxcywh)


INT_BOXES = [[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]]
FLOAT_BOXES = [
[285.3538, 185.5758, 1193.5110, 851.4551],
[285.1472, 188.7374, 1192.4984, 851.0669],
[279.2440, 197.9812, 1189.4746, 849.2019],
]


class TestBoxArea:
def area_check(self, box, expected, atol=1e-4):
out = ops.box_area(box)
Expand Down Expand Up @@ -1152,95 +1144,116 @@ def test_box_area_jit(self):
torch.testing.assert_close(scripted_area, expected)


INT_BOXES = [[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300], [0, 0, 25, 25]]
INT_BOXES2 = [[0, 0, 100, 100], [0, 0, 50, 50], [200, 200, 300, 300]]
FLOAT_BOXES = [
[285.3538, 185.5758, 1193.5110, 851.4551],
[285.1472, 188.7374, 1192.4984, 851.0669],
[279.2440, 197.9812, 1189.4746, 849.2019],
]


class TestIouBase:
@staticmethod
def _run_test(target_fn: Callable, test_input: List, dtypes: List[torch.dtype], atol: float, expected: List):
def _run_test(target_fn: Callable, actual_box1, actual_box2, dtypes, atol, expected):
for dtype in dtypes:
actual_box = torch.tensor(test_input, dtype=dtype)
actual_box1 = torch.tensor(actual_box1, dtype=dtype)
actual_box2 = torch.tensor(actual_box2, dtype=dtype)
expected_box = torch.tensor(expected)
out = target_fn(actual_box, actual_box)
out = target_fn(actual_box1, actual_box2)
print(out)
torch.testing.assert_close(out, expected_box, rtol=0.0, check_dtype=False, atol=atol)

@staticmethod
def _run_jit_test(target_fn: Callable, test_input: List):
box_tensor = torch.tensor(test_input, dtype=torch.float)
def _run_jit_test(target_fn: Callable, actual_box1: List):
box_tensor = torch.tensor(actual_box1, dtype=torch.float)
expected = target_fn(box_tensor, box_tensor)
scripted_fn = torch.jit.script(target_fn)
scripted_out = scripted_fn(box_tensor, box_tensor)
torch.testing.assert_close(scripted_out, expected)


class TestBoxIou(TestIouBase):
int_expected = [[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0]]
int_expected = [[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0], [0.0625, 0.25, 0.0]]
float_expected = [[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]]

@pytest.mark.parametrize(
"test_input, dtypes, atol, expected",
"actual_box1, actual_box2, dtypes, atol, expected",
[
pytest.param(INT_BOXES, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
pytest.param(INT_BOXES, INT_BOXES2, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
],
)
def test_iou(self, test_input, dtypes, atol, expected):
self._run_test(ops.box_iou, test_input, dtypes, atol, expected)
def test_iou(self, actual_box1, actual_box2, dtypes, atol, expected):
self._run_test(ops.box_iou, actual_box1, actual_box2, dtypes, atol, expected)

def test_iou_jit(self):
self._run_jit_test(ops.box_iou, INT_BOXES)


class TestGeneralizedBoxIou(TestIouBase):
int_expected = [[1.0, 0.25, -0.7778], [0.25, 1.0, -0.8611], [-0.7778, -0.8611, 1.0]]
int_expected = [[1.0, 0.25, -0.7778], [0.25, 1.0, -0.8611], [-0.7778, -0.8611, 1.0], [0.0625, 0.25, -0.8819]]
float_expected = [[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]]

@pytest.mark.parametrize(
"test_input, dtypes, atol, expected",
"actual_box1, actual_box2, dtypes, atol, expected",
[
pytest.param(INT_BOXES, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
pytest.param(INT_BOXES, INT_BOXES2, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
],
)
def test_iou(self, test_input, dtypes, atol, expected):
self._run_test(ops.generalized_box_iou, test_input, dtypes, atol, expected)
def test_iou(self, actual_box1, actual_box2, dtypes, atol, expected):
self._run_test(ops.generalized_box_iou, actual_box1, actual_box2, dtypes, atol, expected)

def test_iou_jit(self):
self._run_jit_test(ops.generalized_box_iou, INT_BOXES)


class TestDistanceBoxIoU(TestIouBase):
int_expected = [[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0]]
int_expected = [
[1.0000, 0.1875, -0.4444],
[0.1875, 1.0000, -0.5625],
[-0.4444, -0.5625, 1.0000],
[-0.0781, 0.1875, -0.6267],
]
float_expected = [[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]]

@pytest.mark.parametrize(
"test_input, dtypes, atol, expected",
"actual_box1, actual_box2, dtypes, atol, expected",
[
pytest.param(INT_BOXES, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
pytest.param(INT_BOXES, INT_BOXES2, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
],
)
def test_iou(self, test_input, dtypes, atol, expected):
self._run_test(ops.distance_box_iou, test_input, dtypes, atol, expected)
def test_iou(self, actual_box1, actual_box2, dtypes, atol, expected):
self._run_test(ops.distance_box_iou, actual_box1, actual_box2, dtypes, atol, expected)

def test_iou_jit(self):
self._run_jit_test(ops.distance_box_iou, INT_BOXES)


class TestCompleteBoxIou(TestIouBase):
int_expected = [[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0]]
int_expected = [
[1.0000, 0.1875, -0.4444],
[0.1875, 1.0000, -0.5625],
[-0.4444, -0.5625, 1.0000],
[-0.0781, 0.1875, -0.6267],
]
float_expected = [[1.0, 0.9933, 0.9673], [0.9933, 1.0, 0.9737], [0.9673, 0.9737, 1.0]]

@pytest.mark.parametrize(
"test_input, dtypes, atol, expected",
"actual_box1, actual_box2, dtypes, atol, expected",
[
pytest.param(INT_BOXES, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
pytest.param(INT_BOXES, INT_BOXES2, [torch.int16, torch.int32, torch.int64], 1e-4, int_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float16], 0.002, float_expected),
pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
],
)
def test_iou(self, test_input, dtypes, atol, expected):
self._run_test(ops.complete_box_iou, test_input, dtypes, atol, expected)
def test_iou(self, actual_box1, actual_box2, dtypes, atol, expected):
self._run_test(ops.complete_box_iou, actual_box1, actual_box2, dtypes, atol, expected)

def test_iou_jit(self):
self._run_jit_test(ops.complete_box_iou, INT_BOXES)
Expand Down
12 changes: 7 additions & 5 deletions torchvision/ops/boxes.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -325,13 +325,13 @@ def complete_box_iou(boxes1: Tensor, boxes2: Tensor, eps: float = 1e-7) -> Tenso

diou, iou = _box_diou_iou(boxes1, boxes2, eps)

w_pred = boxes1[:, 2] - boxes1[:, 0]
h_pred = boxes1[:, 3] - boxes1[:, 1]
w_pred = boxes1[:, None, 2] - boxes1[:, None, 0]
Copy link
Contributor

@jdsgomes jdsgomes Aug 2, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why this is needed?

Copy link
Contributor

@datumbox datumbox Aug 2, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is necessary for the broadcasting to work properly. See also:

vision/torchvision/ops/boxes.py

Lines 368 to 371 in ea0be26

lti = torch.min(boxes1[:, None, :2], boxes2[:, :2])
rbi = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])
whi = _upcast(rbi - lti).clamp(min=0) # [N,M,2]
diagonal_distance_squared = (whi[:, :, 0] ** 2) + (whi[:, :, 1] ** 2) + eps

h_pred = boxes1[:, None, 3] - boxes1[:, None, 1]

w_gt = boxes2[:, 2] - boxes2[:, 0]
h_gt = boxes2[:, 3] - boxes2[:, 1]

v = (4 / (torch.pi**2)) * torch.pow((torch.atan(w_gt / h_gt) - torch.atan(w_pred / h_pred)), 2)
v = (4 / (torch.pi**2)) * torch.pow(torch.atan(w_pred / h_pred) - torch.atan(w_gt / h_gt), 2)
with torch.no_grad():
alpha = v / (1 - iou + v + eps)
return diou - alpha * v
Expand All @@ -358,7 +358,7 @@ def distance_box_iou(boxes1: Tensor, boxes2: Tensor, eps: float = 1e-7) -> Tenso

boxes1 = _upcast(boxes1)
boxes2 = _upcast(boxes2)
diou, _ = _box_diou_iou(boxes1, boxes2)
diou, _ = _box_diou_iou(boxes1, boxes2, eps=eps)
return diou


Expand All @@ -375,7 +375,9 @@ def _box_diou_iou(boxes1: Tensor, boxes2: Tensor, eps: float = 1e-7) -> Tuple[Te
x_g = (boxes2[:, 0] + boxes2[:, 2]) / 2
y_g = (boxes2[:, 1] + boxes2[:, 3]) / 2
# The distance between boxes' centers squared.
centers_distance_squared = (_upcast(x_p - x_g) ** 2) + (_upcast(y_p - y_g) ** 2)
centers_distance_squared = (_upcast((x_p[:, None] - x_g[None, :])) ** 2) + (
_upcast((y_p[:, None] - y_g[None, :])) ** 2
)
# The distance IoU is the IoU penalized by a normalized
# distance between boxes' centers squared.
return iou - (centers_distance_squared / diagonal_distance_squared), iou
Expand Down
4 changes: 2 additions & 2 deletions torchvision/ops/ciou_loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@ def complete_box_iou_loss(

"""
Gradient-friendly IoU loss with an additional penalty that is non-zero when the
boxes do not overlap overlap area, This loss function considers important geometrical
boxes do not overlap area, This loss function considers important geometrical
factors such as overlap area, normalized central point distance and aspect ratio.
This loss is symmetric, so the boxes1 and boxes2 arguments are interchangeable.

Expand All @@ -35,7 +35,7 @@ def complete_box_iou_loss(
Tensor: Loss tensor with the reduction option applied.

Reference:
Zhaohui Zheng et. al: Complete Intersection over Union Loss:
Zhaohui Zheng et al.: Complete Intersection over Union Loss:
https://arxiv.org/abs/1911.08287

"""
Expand Down