cameravision

CameraVision

Represent, manipulate and use camera calibration info in computer vision tasks.

Main features:

• The library supports converting coordinates between world, camera and image space, handing lens distortion models according to the Brown–Conrady and Kannala–Brandt models.

• Modify cameras with intuitive methods such as camera.zoom, camera.rotate, camera.scale_output, camera.turn_towards, etc.

• Conversion between distorted and undistorted image spaces are also implemented in an efficient way using Numba and a more accurate inversion of Brown–Conrady distortion compared to OpenCV. We use Newton's method in addition to the standard fixed-point iteration. This library can also keep track of valid image regions after warping, inspired by Leotta et al., but extended to the full Brown-Conrady and Kannala-Brandt models.

• This library also includes efficient implementations of image warping, with antialiasing support and interpolation in linear sRGB color space. The warping maps can be cached for very fast repeated use (e.g., warp/undistort a video taken from a static camera to another calibration setup). This also supports partial caching of only the more expensive distortion part. This is useful when the rotation can change during a video, but the distortion parameters are fixed (e.g., turning the camera to keep the subject centered).

Installation

pip install cameravision

It is recommended to then run the Numba precompilation step (takes around 1–2 minutes). This will make image warping and coordinate transformations fast already on first use.

python -m cameravision.precompile

Documentation

TODO

References

For the idea of computing the valid image region after distortion, see:

• Matthew J. Leotta, David Russell, Andrew Matrai, "On the Maximum Radius of Polynomial Lens Distortion", WACV 2022.