Evaluate camera parameters by seeing if they make image correspondences overlap.
calibration/evaluate_calibration dataset_parameters.json cors.json cams.json out_samples.txt [random/all] [random_count=100000]
Must be given image correspondences with feature point depths. Does not use the images or depth maps in the dataset.
For pairs of view indices (ref,tg), warps feature points of ref to tg, using the extrinsic matrices of the two views, and the feature point depths of ref. Then measures the reprojection error, i.e. the root-mean square distance of the warped feature points and the corresponding feature points from tg. Assuming that the image correspondences are without error, then this error metric only measures the error in the camera parameters.
Can either be run for each possible pair of views (option all). If the dataset if too large for this, can instead be set to sample randomly selected pairs of views (option random).
Outputs out_samples.txt, which contains, for each test, the measured reprojection error, and the baseline of the view indices pair. The baseline is the 2D euclidian distance between view indices.