Deforming Objects Provide Better Camera Calibration

We present a method to calibrate perspective cameras from views of a deforming textured object. We proceed by finding an orthographic calibration, then enriching the camera model to include perspective effects and distortion terms. In the process, we establish the utility of using surface normals to calibrate cameras in the orthographic setting, with a proof that a metric reconstruction can be achieved in two orthographic views of two points combined with two normals. Our calibration object is a sheet of textured cloth. We show that a calibration applies to a specific region of space. Substantial improvements in reconstruction quality and in reprojection error can be obtained by using calibration objects that explore most of the relevant 3D volume. Using cloth as a calibration object allows easy calibration of large 3D volumes more accurately than typical planar calibration objects.