Accuracy

When 3DLDF is run, it uses the three-dimensional data contained in the user code to create a two-dimensional projection. Currently, this can be a perspective projection, or a parallel projection onto one of the major planes. MetaPost code representing this projection is then written to the output file. 3DLDF does no scan conversion,¹ so all of the curves in the projection are generated by means of the algorithms MetaPost inherited from Metafont. These algorithms, however, are designed to find the "most pleasing curve"² given one or more two-dimensional points and connectors; they do not account for the the fact that the two-dimensional points are projections of three-dimensional ones. This can lead to unsatisfactory results, especially where extreme foreshortening occurs. In particular, curl, dir, tension, and control points should be used cautiously, or avoided altogether, when specifying connectors.

3DLDF operates on the assumption that, given an adequate number of points, MetaPost will produce an adequate approximation to the desired curve in perspective, since the greater the number of points given for a curve, the less "choice" MetaPost has for the path through them. My experience with 3DLDF bears this out. Generally, the curves look quite good. Where problems arise, it usually helps to increase the number of points in a curve.

A more serious problem is the imprecision resulting from the operation of rotation. Rotations use the trigonometric functions, which return approximate values. This has the result that points that should have identical coordinate values, sometimes do not. This has consequences for the functions that compare points. The more rotations are applied to points, the greater the divergence between their actual coordinate values, and the values they should have. So far, I haven't found a solution for this problem. On the other hand, it hasn't yet affected the usability of 3DLDF.