We describe a new algorithm for computing the camera heading from
a multi--frame sequence of tracked points, where the camera translates in a
constant direction with arbitrary and possibly varying rotations and speed.
We also adapt our previous general--motion algorithm to recover the complete
structure and motion. An important part of our approach is a new,
reduced--bias linear--subspace method for heading recovery. We point out
that all linear--subspace algorithms have difficulties with planar scenes
and in accurately recovering the focus of expansion when it is outside the
image region, even for nonplanar scenes. We experimentally compare our
approach to a standard two--frame algorithm and to the optimal
least--squares estimate. We demonstrate that accounting for the recently
analyzed ``flipping'' ambiguity improves the robustness of two--frame
reconstruction as well as that of our approach. Our experiments on heading
recovery for planar scenes show that the error landscape has few significant
local minima for such scenes. Our results suggest that the least--squares
error surface for structure from motion may have many local minima when the
focus of expansion lies within the image region.