memorydump85 / undistort

Locally-weighted homography models of lens distortion

undistort

Computes an undistortion table using locally-weighted homographies. The method used is described in:

Locally-weighted Homographies for Calibration of Imaging Systems
Pradeep Ranganathan and Edwin Olson

Quickstart with Docker

Use the provided Dockerfile to build a docker image and run this code inside it.

docker build -t undistort .
docker run --rm -it \
    undistort /code/undistort.py --scale=1.5 /code/examples/example.png

The command above writes the output files to the container's internal filesystem, which is not easily accessible from the host. One way to work around this is to mount the host examples folder into the container.

docker run --rm -it                       \
    -v $PWD/examples/:/code/examples/     \
    undistort /code/undistort.py --scale=1.5 /code/examples/example.png

This command runs the undistortion script on the examples/example.png and produces the undistortion table examples/example.png.table and rectified image examples/example.fixed.png.

The produced .table file can be used to undistort other images using the examples/render_undistorted.py script.

Notes

The file examples/example.png is an example of an ideal image for estimating distortion. It provides features that can cover the entire image. The non-parametric nature of the distortion model makes it data-intensive -- the model is accurate near observed data points. Hence an image with detections that covers the entire image produces better models.

The file examples/bad_example.png is an example of an image that might not produce a good distortion model.