Skip to content

Commit

Permalink
Conversion from OpenCV cameras
Browse files Browse the repository at this point in the history
Summary: Implements a conversion function between OpenCV and PyTorch3D cameras.

Reviewed By: patricklabatut

Differential Revision: D28992470

fbshipit-source-id: dbcc9f213ec293c2f6938261c704aea09aad3c90
  • Loading branch information
davnov134 authored and facebook-github-bot committed Jun 21, 2021
1 parent b2ac265 commit 8006842
Show file tree
Hide file tree
Showing 4 changed files with 1,450 additions and 0 deletions.
1 change: 1 addition & 0 deletions pytorch3d/utils/__init__.py
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.

from .camera_conversions import cameras_from_opencv_projection
from .ico_sphere import ico_sphere
from .torus import torus

Expand Down
70 changes: 70 additions & 0 deletions pytorch3d/utils/camera_conversions.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,70 @@
import torch

from ..renderer import PerspectiveCameras
from ..transforms import so3_exponential_map


def cameras_from_opencv_projection(
rvec: torch.Tensor,
tvec: torch.Tensor,
camera_matrix: torch.Tensor,
image_size: torch.Tensor,
) -> PerspectiveCameras:
"""
Converts a batch of OpenCV-conventioned cameras parametrized with the
axis-angle rotation vectors `rvec`, translation vectors `tvec`, and the camera
calibration matrices `camera_matrix` to `PerspectiveCameras` in PyTorch3D
convention.
More specifically, the conversion is carried out such that a projection
of a 3D shape to the OpenCV-conventioned screen of size `image_size` results
in the same image as a projection with the corresponding PyTorch3D camera
to the NDC screen convention of PyTorch3D.
More specifically, the OpenCV convention projects points to the OpenCV screen
space as follows:
```
x_screen_opencv = camera_matrix @ (exp(rvec) @ x_world + tvec)
```
followed by the homogenization of `x_screen_opencv`.
Note:
The parameters `rvec, tvec, camera_matrix` correspond e.g. to the inputs
of `cv2.projectPoints`, or to the ouputs of `cv2.calibrateCamera`.
Args:
rvec: A batch of axis-angle rotation vectors of shape `(N, 3)`.
tvec: A batch of translation vectors of shape `(N, 3)`.
camera_matrix: A batch of camera calibration matrices of shape `(N, 3, 3)`.
image_size: A tensor of shape `(N, 2)` containing the sizes of the images
(height, width) attached to each camera.
Returns:
cameras_pytorch3d: A batch of `N` cameras in the PyTorch3D convention.
"""

R = so3_exponential_map(rvec)
focal_length = torch.stack([camera_matrix[:, 0, 0], camera_matrix[:, 1, 1]], dim=-1)
principal_point = camera_matrix[:, :2, 2]

# Retype the image_size correctly and flip to width, height.
image_size_wh = image_size.to(R).flip(dims=(1,))

# Get the PyTorch3D focal length and principal point.
focal_pytorch3d = focal_length / (0.5 * image_size_wh)
p0_pytorch3d = -(principal_point / (0.5 * image_size_wh) - 1)

# For R, T we flip x, y axes (opencv screen space has an opposite
# orientation of screen axes).
# We also transpose R (opencv multiplies points from the opposite=left side).
R_pytorch3d = R.permute(0, 2, 1)
T_pytorch3d = tvec.clone()
R_pytorch3d[:, :, :2] *= -1
T_pytorch3d[:, :2] *= -1

return PerspectiveCameras(
R=R_pytorch3d,
T=T_pytorch3d,
focal_length=focal_pytorch3d,
principal_point=p0_pytorch3d,
)
Loading

0 comments on commit 8006842

Please sign in to comment.