Skip to content
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

[proto] Small optims for perspective bboxes op #6891

Merged
merged 14 commits into from
Nov 3, 2022
Merged

[proto] Small optims for perspective bboxes op #6891

Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
6a618e2
[proto] Speed-up crop on bboxes and tests
vfdev-5 Nov 1, 2022
71e952d
Fix linter
vfdev-5 Nov 1, 2022
9baed3a
Update _geometry.py
vfdev-5 Nov 1, 2022
9abc115
Merge branch 'main' into proto-speedup-crop-bboxes
vfdev-5 Nov 1, 2022
0b11d0b
Fixed device issue
vfdev-5 Nov 1, 2022
04b893b
Revert changes in test/prototype_transforms_kernel_infos.py
vfdev-5 Nov 1, 2022
41abd73
Fixed failing correctness tests
vfdev-5 Nov 2, 2022
2fe32df
Merge branch 'proto-speedup-crop-bboxes' into proto-speedup-pad-bboxes
vfdev-5 Nov 2, 2022
56f880a
[proto] Optimized functional pad op for bboxes + tests
vfdev-5 Nov 2, 2022
18f40f1
Renamed copy-pasted variable name
vfdev-5 Nov 2, 2022
61263ca
[proto] Small optims for perspective bboxes op
vfdev-5 Nov 2, 2022
fa94643
Merge branch 'main' into proto-update-perspective-bboxes
vfdev-5 Nov 2, 2022
1435b42
Merge branch 'main' into proto-update-perspective-bboxes
vfdev-5 Nov 3, 2022
2c372a9
Merge branch 'main' into proto-update-perspective-bboxes
vfdev-5 Nov 3, 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
23 changes: 12 additions & 11 deletions torchvision/prototype/transforms/functional/_geometry.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -928,16 +928,16 @@ def perspective_bounding_box(
(-perspective_coeffs[0] * perspective_coeffs[7] + perspective_coeffs[1] * perspective_coeffs[6]) / denom,
]

theta1 = torch.tensor(
[[inv_coeffs[0], inv_coeffs[1], inv_coeffs[2]], [inv_coeffs[3], inv_coeffs[4], inv_coeffs[5]]],
theta12_T = torch.tensor(
[
[inv_coeffs[0], inv_coeffs[3], inv_coeffs[6], inv_coeffs[6]],
[inv_coeffs[1], inv_coeffs[4], inv_coeffs[7], inv_coeffs[7]],
[inv_coeffs[2], inv_coeffs[5], 1.0, 1.0],
],
dtype=dtype,
device=device,
)

theta2 = torch.tensor(
[[inv_coeffs[6], inv_coeffs[7], 1.0], [inv_coeffs[6], inv_coeffs[7], 1.0]], dtype=dtype, device=device
)

# 1) Let's transform bboxes into a tensor of 4 points (top-left, top-right, bottom-left, bottom-right corners).
# Tensor of points has shape (N * 4, 3), where N is the number of bboxes
# Single point structure is similar to
Expand All @@ -948,15 +948,16 @@ def perspective_bounding_box(
# x_out = (coeffs[0] * x + coeffs[1] * y + coeffs[2]) / (coeffs[6] * x + coeffs[7] * y + 1)
# y_out = (coeffs[3] * x + coeffs[4] * y + coeffs[5]) / (coeffs[6] * x + coeffs[7] * y + 1)

numer_points = torch.matmul(points, theta1.T)
denom_points = torch.matmul(points, theta2.T)
transformed_points = numer_points / denom_points
numer_denom_points = torch.matmul(points, theta12_T)
numer_points = numer_denom_points[:, :2]
denom_points = numer_denom_points[:, 2:]
transformed_points = numer_points.div_(denom_points)

# 3) Reshape transformed points to [N boxes, 4 points, x/y coords]
# and compute bounding box from 4 transformed points:
transformed_points = transformed_points.reshape(-1, 4, 2)
out_bbox_mins, _ = torch.min(transformed_points, dim=1)
out_bbox_maxs, _ = torch.max(transformed_points, dim=1)
out_bbox_mins, out_bbox_maxs = torch.aminmax(transformed_points, dim=1)

out_bboxes = torch.cat([out_bbox_mins, out_bbox_maxs], dim=1).to(bounding_box.dtype)

# out_bboxes should be of shape [N boxes, 4]
Expand Down