CamTools: Camera Tools for Computer Vision.
Project description
CamTools: Camera Tools for Computer Vision
CamTools is a useful tool for handling cameras in computer vision. It can be used for plotting, converting, projecting, ray casting, and doing more with camera parameters. It follows the standard camera coordinate system with clear and easy-to-use APIs.
What can you do with CamTools?
-
Plot cameras. Useful for debugging 3D reconstruction and NeRFs!
import camtools as ct import open3d as o3d cameras = ct.camera.create_camera_ray_frames(Ks, Ts) o3d.visualization.draw_geometries([cameras])
-
Convert camera parameters.
pose = ct.convert.T_to_pose(T) # Convert T to pose R, t = ct.convert.T_to_R_t(T) # Convert T to R and t C = ct.convert.pose_to_C(pose) # Convert pose to camera center K, T = ct.convert.P_to_K_T(P) # Decompose projection matrix to K and T # And more...
-
Projection and ray casting.
# Project 3D points to pixels. pixels = ct.project.points_to_pixel(points, K, T) # Back-project depth image to 3D points. points = ct.project.im_depth_to_points(im_depth, K, T) # Ray cast a triangle mesh to depth image. im_depth = ct.raycast.mesh_to_depths(mesh, Ks, Ts, height, width) # And more...
-
Image and depth I/O with no surprises.
Strict type checks and range checks are enforced. The image and depth I/O APIs are specifically designed to solve the following pain points:
- Is my image of type
float32
oruint8
? - Does it have range
[0, 1]
or[0, 255]
? - Is it RGB or BGR?
- Does my image have an alpha channel?
- When saving depth image as integer-based
.png
, is it correctly scaled?
ct.io.imread() ct.io.imwrite() ct.io.imread_detph() ct.io.imwrite_depth()
- Is my image of type
-
Command-line tools
ct
(runs in terminal).# Crop image boarders. ct crop-boarders *.png --pad_pixel 10 --skip_cropped --same_crop # Draw synchronized bounding boxes interactively. ct draw-bboxes path/to/a.png path/to/b.png # For more command-line tools. ct --help
-
And more.
- Solve line intersections.
- COLMAP tools.
- Points normalization.
- ...
Installation
# Option 1: install from pip.
pip install camtools
# Option 2: install from git.
pip install git+https://github.com/yxlao/camtools.git
# Option 3: install from source.
git clone https://github.com/yxlao/camtools.git
cd camtools
pip install -e . # Dev mode, if you want to modify camtools.
pip install . # Install mode, if you want to use camtools only.
Camera coordinate system
We follow the standard OpenCV-style camera coordinate system as shown below.
- Camera coordinate: right-handed, with $Z$ pointing away from the camera towards the view direction and $Y$ axis pointing down. Note that this is different from the Blender convention, where $Z$ points towards the opposite view direction and the $Y$ axis points up.
- Image coordinate: starts from the top-left corner of the image, with $x$
pointing right (corresponding to the image width) and $y$ pointing down
(corresponding to the image height). This is consistent with OpenCV. Pay
attention that the 0th dimension in the image array is the height (i.e., $y$)
and the 1st dimension is the width (i.e., $x$). That is:
- $x$ <=> $u$ <=> width <=> column <=> the 1st dimension
- $y$ <=> $v$ <=> height <=> row <=> the 0th dimension
K
:(3, 3)
camera intrinsic matrix.K = [[fx, s, cx], [ 0, fy, cy], [ 0, 0, 1]]
T
orW2C
:(4, 4)
camera extrinsic matrix.T = [[R | t = [[R_01, R_02, R_03, t_0], 0 | 1]] [R_11, R_12, R_13, t_1], [R_21, R_22, R_23, t_2], [ 0, 0, 0, 1]]
T
is also known as the world-to-cameraW2C
matrix, which transforms a point in the world coordinate to the camera coordinate.T
's shape is(4, 4)
, not(3, 4)
.T
must be invertible, wherenp.linalg.inv(T) = pose
.- The camera center
C
in world coordinate is projected to[0, 0, 0, 1]
in camera coordinate, i.e.,T @ C = np.array([0, 0, 0, 1]).T
R
:(3, 3)
rotation matrix.R = T[:3, :3]
R
is a rotation matrix. It is an orthogonal matrix with determinant 1, as rotations preserve volume and orientation.R.T == np.linalg.inv(R)
np.linalg.norm(R @ x) == np.linalg.norm(x)
, wherex
is a(3,)
vector.
t
:(3,)
translation vector.t = T[:3, 3]
t
's shape is(3,)
, not(3, 1)
.
pose
orC2W
:(4, 4)
camera pose matrix. It is the inverse ofT
.pose = T.inv()
pose
is also known as the camera-to-worldC2W
matrix, which transforms a point in the camera coordinate to the world coordinate.pose
is the inverse ofT
, i.e.,pose == np.linalg.inv(T)
.
C
: camera center.C = pose[:3, 3]
C
's shape is(3,)
, not(3, 1)
.C
is the camera center in world coordinate. It is also the translation vector ofpose
.
P
:(3, 4)
the camera projection matrix.P
is the world-to-pixel projection matrix, which projects a point in the homogeneous world coordinate to the homogeneous pixel coordinate.P
is the product of the intrinsic and extrinsic parameters.# P = K @ [R | t] P = K @ np.hstack([R, t[:, None]])
P
's shape is(3, 4)
, not(4, 4)
.- It is possible to decompose
P
into intrinsic and extrinsic matrices by QR decomposition. - Don't confuse
P
withpose
.
- For more details, please refer to the following blog posts: part 1, part 2, and part 3.
Future works
- Refined APIs.
- Full PyTorch/Numpy compatibility.
- Unit tests.
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.