Systems that track a camera through video are usually graded by absolute trajectory error: align the predicted camera positions to the true ones and measure the gap. A low gap is read as proof the system saw the scene.
To test that, one can score a fake path that uses no images: a straight line at constant speed. They also rotated predicted positions away from the directions the cameras really faced, which should ruin the geometry.
On the Sintel video benchmark, the image-free line beat 10 of 11 recent image-based models. Rotating positions left the score unchanged even though scene points then landed in the wrong place. A revised score that uses camera orientations catches this.