mguti97/PnLCalib ? reverse-engineered prompt

Build me a Python project that can calibrate a soccer broadcast camera from an image or video of the field.

It should detect field keypoints and field lines, then use them to estimate the camera pose or homography, with an optional refinement step that improves the result by combining points and lines. I want to be able to run it on a single image or a video, save the output with the projected field overlay, and optionally display the result live.

Please include training and inference scripts, plus a simple way to run evaluation on common soccer calibration datasets like SoccerNet Calibration, WorldCup 2014, and TS WorldCup. Use pretrained weights for the keypoint and line models if available, and make it easy to swap in different weights for single view or multi view setups.

If you need to check current dataset or dependency details online, feel free to look them up.