A variant of the trackball interaction works like this: The initial click is at some point P of the image plane; the mouse is currently at some point Q. The center of the object is at C, which is assumed to not be on the image plane. The vector (Q - C)×(P - C) serves as an axis for the rotation, with the amount of rotation made proportional to Q - P. It's nice if the proportionality constant is chosen so that for small drags, at least those that start on the line between the eyepoint and C, the rotation resembles the rotation provided by the virtual sphere interaction. The advantage is that there's no special-case handling needed if the drag goes off the transparent sphere. Implement it to see if there are any obvious disadvantages. Can you easily spin the object around the eye-to-object axis?