def test_transform_boxes_wh(self, device, dtype):
boxes = torch.tensor(
[
[139.2640, 103.0150, 258.0480, 307.5075],
[1.0240, 80.5547, 510.9760, 431.4453],
[165.2053, 262.1440, 345.4293, 246.7840],
[119.8080, 144.2067, 137.2160, 265.9225],
],
device=device,
dtype=dtype,
)
expected = torch.tensor(
[
[372.7360, 103.0150, -258.0480, 307.5075],
[510.9760, 80.5547, -510.9760, 431.4453],
[346.7947, 262.1440, -345.4293, 246.7840],
[392.1920, 144.2067, -137.2160, 265.9225],
],
device=device,
dtype=dtype,
)
trans_mat = torch.tensor([[[-1.0, 0.0, 512.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]], device=device, dtype=dtype)
out = transform_bbox(trans_mat, boxes, mode='xywh')
assert_allclose(out, expected, atol=1e-4, rtol=1e-4)
def test_transform_boxes_wh(self, device, dtype):
boxes = torch.tensor(
[
[139.2640, 103.0150, 258.0480, 307.5075],
[1.0240, 80.5547, 510.9760, 431.4453],
[165.2053, 262.1440, 345.4293, 246.7840],
[119.8080, 144.2067, 137.2160, 265.9225],
],
device=device,
dtype=dtype,
)
expected = torch.tensor(
[
[114.6880, 103.0150, 258.0480, 307.5075],
[0.0000, 80.5547, 510.9760, 431.4453],
[1.3654, 262.1440, 345.4293, 246.7840],
[254.9760, 144.2067, 137.2160, 265.9225]
],
device=device,
dtype=dtype,
)
trans_mat = torch.tensor([[[-1.0, 0.0, 512.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]], device=device, dtype=dtype)
out = transform_bbox(trans_mat, boxes, mode='xywh', restore_coordinates=True)
assert_allclose(out, expected, atol=1e-4, rtol=1e-4)
def test_transform_boxes(self, device, dtype):
boxes = torch.tensor([[139.2640, 103.0150, 397.3120, 410.5225]], device=device, dtype=dtype)
expected = torch.tensor([[372.7360, 103.0150, 114.6880, 410.5225]], device=device, dtype=dtype)
trans_mat = torch.tensor([[[-1.0, 0.0, 512.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]], device=device, dtype=dtype)
out = transform_bbox(trans_mat, boxes)
assert_allclose(out, expected, atol=1e-4, rtol=1e-4)
def inverse_bbox(self,
input: torch.Tensor,
module: nn.Module,
param: Optional[ParamItem] = None,
mode: str = "xyxy") -> torch.Tensor:
if isinstance(module, GeometricAugmentationBase2D):
transform = module.compute_inverse_transformation(
module.get_transformation_matrix(
input, None if param is None else cast(Dict, param.data)))
input = transform_bbox(
torch.as_tensor(transform,
device=input.device,
dtype=input.dtype), input, mode)
return input
def test_transform_multiple_boxes(self, device, dtype):
boxes = torch.tensor(
[
[139.2640, 103.0150, 397.3120, 410.5225],
[1.0240, 80.5547, 512.0000, 512.0000],
[165.2053, 262.1440, 510.6347, 508.9280],
[119.8080, 144.2067, 257.0240, 410.1292],
],
device=device,
dtype=dtype,
)
boxes = boxes.repeat(2, 1, 1) # 2 x 4 x 4 two images 4 boxes each
expected = torch.tensor(
[
[
[372.7360, 103.0150, 114.6880, 410.5225],
[510.9760, 80.5547, 0.0000, 512.0000],
[346.7947, 262.1440, 1.3653, 508.9280],
[392.1920, 144.2067, 254.9760, 410.1292],
],
[
[139.2640, 103.0150, 397.3120, 410.5225],
[1.0240, 80.5547, 512.0000, 512.0000],
[165.2053, 262.1440, 510.6347, 508.9280],
[119.8080, 144.2067, 257.0240, 410.1292],
],
],
device=device,
dtype=dtype,
)
trans_mat = torch.tensor(
[
[[-1.0, 0.0, 512.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]],
[[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]],
],
device=device,
dtype=dtype,
)
out = transform_bbox(trans_mat, boxes)
assert_allclose(out, expected, atol=1e-4, rtol=1e-4)
def apply_to_bbox(self,
input: torch.Tensor,
module: nn.Module,
param: Optional[ParamItem] = None,
mode: str = "xyxy") -> torch.Tensor:
if param is not None:
_param = cast(Dict[str, torch.Tensor], param.data)
else:
_param = None # type: ignore
if isinstance(module, GeometricAugmentationBase2D) and _param is None:
raise ValueError(
f"Transformation matrix for {module} has not been computed.")
if isinstance(module,
GeometricAugmentationBase2D) and _param is not None:
input = transform_bbox(
module.get_transformation_matrix(input, _param), input, mode)
else:
pass # No need to update anything
return input
