def test_pairwise_iou_many_boxes(self):
for device in ["cpu"
] + (["cuda"] if torch.cuda.is_available() else []):
num_boxes1 = 100
num_boxes2 = 200
boxes1 = torch.stack([
torch.tensor(
[5 + 20 * i, 5 + 20 * i, 10, 10, 0],
dtype=torch.float32,
device=device,
) for i in range(num_boxes1)
])
boxes2 = torch.stack([
torch.tensor(
[5 + 20 * i, 5 + 20 * i, 10, 1 + 9 * i / num_boxes2, 0],
dtype=torch.float32,
device=device,
) for i in range(num_boxes2)
])
expected_ious = torch.zeros(num_boxes1,
num_boxes2,
dtype=torch.float32,
device=device)
for i in range(min(num_boxes1, num_boxes2)):
expected_ious[i][i] = (1 + 9 * i / num_boxes2) / 10.0
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_0_degree_cuda(self):
device = torch.device("cuda")
boxes1 = torch.tensor(
[[0.5, 0.5, 1.0, 1.0, 0.0], [0.5, 0.5, 1.0, 1.0, 0.0]],
dtype=torch.float32,
device=device,
)
boxes2 = torch.tensor(
[
[0.5, 0.5, 1.0, 1.0, 0.0],
[0.25, 0.5, 0.5, 1.0, 0.0],
[0.5, 0.25, 1.0, 0.5, 0.0],
[0.25, 0.25, 0.5, 0.5, 0.0],
[0.75, 0.75, 0.5, 0.5, 0.0],
[1.0, 1.0, 1.0, 1.0, 0.0],
],
dtype=torch.float32,
device=device,
)
expected_ious = torch.tensor(
[
[1.0, 0.5, 0.5, 0.25, 0.25, 0.25 / (2 - 0.25)],
[1.0, 0.5, 0.5, 0.25, 0.25, 0.25 / (2 - 0.25)],
],
dtype=torch.float32,
device=device,
)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_normalize_angles(self):
# torch.manual_seed(0)
for _ in range(50):
num_boxes = 100
boxes_5d = torch.zeros(num_boxes, 5)
boxes_5d[:, 0] = torch.FloatTensor(num_boxes).uniform_(-100, 500)
boxes_5d[:, 1] = torch.FloatTensor(num_boxes).uniform_(-100, 500)
boxes_5d[:, 2] = torch.FloatTensor(num_boxes).uniform_(0, 500)
boxes_5d[:, 3] = torch.FloatTensor(num_boxes).uniform_(0, 500)
boxes_5d[:, 4] = torch.FloatTensor(num_boxes).uniform_(-1800, 1800)
rotated_boxes = RotatedBoxes(boxes_5d)
normalized_boxes = rotated_boxes.clone()
normalized_boxes.normalize_angles()
self.assertTrue(torch.all(normalized_boxes.tensor[:, 4] >= -180))
self.assertTrue(torch.all(normalized_boxes.tensor[:, 4] < 180))
# x, y, w, h should not change
self.assertTrue(torch.allclose(boxes_5d[:, :4], normalized_boxes.tensor[:, :4]))
# the cos/sin values of the angles should stay the same
self.assertTrue(
torch.allclose(
torch.cos(boxes_5d[:, 4] * math.pi / 180),
torch.cos(normalized_boxes.tensor[:, 4] * math.pi / 180),
atol=1e-5,
)
)
self.assertTrue(
torch.allclose(
torch.sin(boxes_5d[:, 4] * math.pi / 180),
torch.sin(normalized_boxes.tensor[:, 4] * math.pi / 180),
atol=1e-5,
)
)
def test_pairwise_iou_0_degree(self):
for device in ["cpu"
] + (["cuda"] if torch.cuda.is_available() else []):
boxes1 = torch.tensor(
[[0.5, 0.5, 1.0, 1.0, 0.0], [0.5, 0.5, 1.0, 1.0, 0.0]],
dtype=torch.float32,
device=device,
)
boxes2 = torch.tensor(
[
[0.5, 0.5, 1.0, 1.0, 0.0],
[0.25, 0.5, 0.5, 1.0, 0.0],
[0.5, 0.25, 1.0, 0.5, 0.0],
[0.25, 0.25, 0.5, 0.5, 0.0],
[0.75, 0.75, 0.5, 0.5, 0.0],
[1.0, 1.0, 1.0, 1.0, 0.0],
],
dtype=torch.float32,
device=device,
)
expected_ious = torch.tensor(
[
[1.0, 0.5, 0.5, 0.25, 0.25, 0.25 / (2 - 0.25)],
[1.0, 0.5, 0.5, 0.25, 0.25, 0.25 / (2 - 0.25)],
],
dtype=torch.float32,
device=device,
)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def func_cat(x: torch.Tensor):
boxes1 = RotatedBoxes(x)
boxes2 = RotatedBoxes(x)
# this is not supported by torchscript for now.
# boxes3 = RotatedBoxes.cat([boxes1, boxes2])
boxes3 = boxes1.cat([boxes1, boxes2])
return boxes3
def test_pairwise_iou_many_boxes_cuda(self):
device = torch.device("cuda")
num_boxes1 = 100
num_boxes2 = 200
boxes1 = torch.stack(
[
torch.tensor(
[5 + 20 * i, 5 + 20 * i, 10, 10, 0], dtype=torch.float32, device=device
)
for i in range(num_boxes1)
]
)
boxes2 = torch.stack(
[
torch.tensor(
[5 + 20 * i, 5 + 20 * i, 10, 1 + 9 * i / num_boxes2, 0],
dtype=torch.float32,
device=device,
)
for i in range(num_boxes2)
]
)
expected_ious = torch.zeros(num_boxes1, num_boxes2, dtype=torch.float32, device=device)
for i in range(min(num_boxes1, num_boxes2)):
expected_ious[i][i] = (1 + 9 * i / num_boxes2) / 10.0
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_orthogonal_cuda(self):
device = torch.device("cuda")
boxes1 = torch.tensor([[5, 5, 10, 6, 55]], dtype=torch.float32, device=device)
boxes2 = torch.tensor([[5, 5, 10, 6, -35]], dtype=torch.float32, device=device)
iou = (6.0 * 6.0) / (6.0 * 6.0 + 4.0 * 6.0 + 4.0 * 6.0)
expected_ious = torch.tensor([[iou]], dtype=torch.float32, device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_issue1207_simplified(self):
for device in ["cpu"] + ["cuda"] if torch.cuda.is_available() else []:
# Simplified test case of D2-issue-1207
boxes1 = torch.tensor([[3, 3, 8, 2, -45.0]], device=device)
boxes2 = torch.tensor([[6, 0, 8, 2, -45.0]], device=device)
iou = 0.0
expected_ious = torch.tensor([[iou]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_issue1207_simplified_cuda(self):
device = torch.device("cuda")
# Simplified test case of D2-issue-1207
boxes1 = torch.tensor([[3, 3, 8, 2, -45.0]], device=device)
boxes2 = torch.tensor([[6, 0, 8, 2, -45.0]], device=device)
iou = 0.0
expected_ious = torch.tensor([[iou]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_45_degrees_cuda(self):
device = torch.device("cuda")
boxes1 = torch.tensor(
[
[1, 1, math.sqrt(2), math.sqrt(2), 45],
[1, 1, 2 * math.sqrt(2), 2 * math.sqrt(2), -45],
],
dtype=torch.float32,
device=device,
)
boxes2 = torch.tensor([[1, 1, 2, 2, 0]], dtype=torch.float32, device=device)
expected_ious = torch.tensor([[0.5], [0.5]], dtype=torch.float32, device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_orthogonal(self):
for device in ["cpu"] + ["cuda"] if torch.cuda.is_available() else []:
boxes1 = torch.tensor([[5, 5, 10, 6, 55]],
dtype=torch.float32,
device=device)
boxes2 = torch.tensor([[5, 5, 10, 6, -35]],
dtype=torch.float32,
device=device)
iou = (6.0 * 6.0) / (6.0 * 6.0 + 4.0 * 6.0 + 4.0 * 6.0)
expected_ious = torch.tensor([[iou]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_issue1207(self):
for device in ["cpu"] + ["cuda"] if torch.cuda.is_available() else []:
# The original test case in D2-issue-1207
boxes1 = torch.tensor([[160.0, 153.0, 230.0, 23.0, -37.0]],
device=device)
boxes2 = torch.tensor([[190.0, 127.0, 80.0, 21.0, -46.0]],
device=device)
iou = 0.0
expected_ious = torch.tensor([[iou]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_issue1207_cuda(self):
device = torch.device("cuda")
# The original test case in D2-issue-1207
boxes1 = torch.tensor([[160.0, 153.0, 230.0, 23.0, -37.0]],
device=device)
boxes2 = torch.tensor([[190.0, 127.0, 80.0, 21.0, -46.0]],
device=device)
iou = 0.0
expected_ious = torch.tensor([[iou]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_large_close_boxes_cuda(self):
device = torch.device("cuda")
boxes1 = torch.tensor(
[[299.500000, 417.370422, 600.000000, 364.259186, 27.1828]],
dtype=torch.float32,
device=device,
)
boxes2 = torch.tensor(
[[299.500000, 417.370422, 600.000000, 364.259155, 27.1828]],
dtype=torch.float32,
device=device,
)
iou = 364.259155 / 364.259186
expected_ious = torch.tensor([[iou]], dtype=torch.float32, device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_large_close_boxes(self):
for device in ["cpu"] + ["cuda"] if torch.cuda.is_available() else []:
boxes1 = torch.tensor(
[[299.500000, 417.370422, 600.000000, 364.259186, 27.1828]],
dtype=torch.float32,
device=device,
)
boxes2 = torch.tensor(
[[299.500000, 417.370422, 600.000000, 364.259155, 27.1828]],
dtype=torch.float32,
device=device,
)
iou = 364.259155 / 364.259186
expected_ious = torch.tensor([[iou]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_pairwise_iou_45_degrees(self):
for device in ["cpu"] + ["cuda"] if torch.cuda.is_available() else []:
boxes1 = torch.tensor(
[
[1, 1, math.sqrt(2), math.sqrt(2), 45],
[1, 1, 2 * math.sqrt(2), 2 * math.sqrt(2), -45],
],
dtype=torch.float32,
device=device,
)
boxes2 = torch.tensor([[1, 1, 2, 2, 0]],
dtype=torch.float32,
device=device)
expected_ious = torch.tensor([[0.5], [0.5]],
dtype=torch.float32,
device=device)
ious = pairwise_iou(RotatedBoxes(boxes1), RotatedBoxes(boxes2))
self.assertTrue(torch.allclose(ious, expected_ious))
def test_clip_area_0_degree(self):
for _ in range(50):
num_boxes = 100
boxes_5d = torch.zeros(num_boxes, 5)
boxes_5d[:, 0] = torch.FloatTensor(num_boxes).uniform_(-100, 500)
boxes_5d[:, 1] = torch.FloatTensor(num_boxes).uniform_(-100, 500)
boxes_5d[:, 2] = torch.FloatTensor(num_boxes).uniform_(0, 500)
boxes_5d[:, 3] = torch.FloatTensor(num_boxes).uniform_(0, 500)
# Convert from (x_ctr, y_ctr, w, h, 0) to (x1, y1, x2, y2)
boxes_4d = torch.zeros(num_boxes, 4)
boxes_4d[:, 0] = boxes_5d[:, 0] - boxes_5d[:, 2] / 2.0
boxes_4d[:, 1] = boxes_5d[:, 1] - boxes_5d[:, 3] / 2.0
boxes_4d[:, 2] = boxes_5d[:, 0] + boxes_5d[:, 2] / 2.0
boxes_4d[:, 3] = boxes_5d[:, 1] + boxes_5d[:, 3] / 2.0
image_size = (500, 600)
test_boxes_4d = Boxes(boxes_4d)
test_boxes_5d = RotatedBoxes(boxes_5d)
# Before clip
areas_4d = test_boxes_4d.area()
areas_5d = test_boxes_5d.area()
self.assertTrue(torch.allclose(areas_4d, areas_5d, atol=1e-1, rtol=1e-5))
# After clip
test_boxes_4d.clip(image_size)
test_boxes_5d.clip(image_size)
areas_4d = test_boxes_4d.area()
areas_5d = test_boxes_5d.area()
self.assertTrue(torch.allclose(areas_4d, areas_5d, atol=1e-1, rtol=1e-5))
def test_clip_area_arbitrary_angle(self):
num_boxes = 100
boxes_5d = torch.zeros(num_boxes, 5)
boxes_5d[:, 0] = torch.FloatTensor(num_boxes).uniform_(-100, 500)
boxes_5d[:, 1] = torch.FloatTensor(num_boxes).uniform_(-100, 500)
boxes_5d[:, 2] = torch.FloatTensor(num_boxes).uniform_(0, 500)
boxes_5d[:, 3] = torch.FloatTensor(num_boxes).uniform_(0, 500)
boxes_5d[:, 4] = torch.FloatTensor(num_boxes).uniform_(-1800, 1800)
clip_angle_threshold = random.uniform(0, 180)
image_size = (500, 600)
test_boxes_5d = RotatedBoxes(boxes_5d)
# Before clip
areas_before = test_boxes_5d.area()
# After clip
test_boxes_5d.clip(image_size, clip_angle_threshold)
areas_diff = test_boxes_5d.area() - areas_before
# the areas should only decrease after clipping
self.assertTrue(torch.all(areas_diff <= 0))
# whenever the box is clipped (thus the area shrinks),
# the angle for the box must be within the clip_angle_threshold
# Note that the clip function will normalize the angle range
# to be within (-180, 180]
self.assertTrue(
torch.all(torch.abs(boxes_5d[:, 4][torch.where(areas_diff < 0)]) < clip_angle_threshold)
)
def test_empty_cat(self):
x = RotatedBoxes.cat([])
self.assertTrue(x.tensor.shape, (0, 5))
def func(x):
boxes = RotatedBoxes(x)
test = boxes.to(torch.device("cpu")).tensor
return boxes.area(), test
def rotated_mapper(original_dataset_dict):
# Implement a mapper, similar to the default DatasetMapper, but with our own customizations
dataset_dict = copy.deepcopy(
original_dataset_dict) # it will be modified by code below
original_gsd = dataset_dict["gsd"]
target_gsd = np.random.uniform(0.09, 0.13) # randomize target gsd
scale = original_gsd / target_gsd
target_size = 400
target_crop = int(target_size / scale)
target_crop = (target_crop, target_crop)
image_np = detection_utils.read_image(dataset_dict["file_name"],
format="BGR")
boxes = np.asarray([anno['bbox'] for anno in dataset_dict['annotations']])
# select anno at random
# draw random center
# h, w = image_np.shape[:2]
# rand_box = boxes[np.random.randint(len(boxes))]
# ch, cw = rand_box[:2]
# xmin = np.min()
# xmax = np.max()
# ymin = 3
# ymax = 4
# h0 = np.random.randint(min(h, ymin), min(h, ymax) + 1)
# w0 = np.random.randint(min(w, xmin), min(w, xmax) + 1)
# assert h >= target_crop[1] and w >= target_crop[0], "Shape computation has bugs."
# crop = T.CropTransform(w0, h0, target_crop)
# make sure random crop contains annotations
i = 0
while True:
random_crop = T.RandomCrop('absolute',
target_crop).get_transform(image_np)
cropped_boxes = RotatedBoxes(
random_crop.apply_coords(copy.deepcopy(boxes)))
inside_ind = cropped_boxes.inside_box(target_crop)
if 1 < sum(inside_ind) <= 100:
break
i += 1
if i > 150:
return None
image, transforms = T.apply_transform_gens([
random_crop,
T.Resize((target_size, target_size)),
], image_np)
dataset_dict["image"] = torch.as_tensor(
image.transpose(2, 0, 1).astype("float32"))
annos = [
rotated_transform_instance_annotations(obj, transforms,
image.shape[:2])
for obj in dataset_dict.pop("annotations")
if obj.get("iscrowd", 0) == 0
]
instances = detection_utils.annotations_to_instances_rotated(
annos, image.shape[:2])
instances = detection_utils.filter_empty_instances(instances)
inside_ind = instances.gt_boxes.inside_box(image.shape[:2])
instances = instances[inside_ind]
assert ((instances.gt_boxes.tensor.numpy()[:, 2] > 0).all().item()
), "width not > 0\n\n" + str(instances.gt_boxes.tensor.numpy())
dataset_dict["instances"] = instances
return dataset_dict
