rs 
Indeed, with strong multistation configurations it will be shown later that 
the inclusion of any object space control, whether related to the object 
targets or camera exterior orientation parameters, is of negligible value. 
Further, if the co-ordinate system chosen is arbitrary (eg. if we are only 
interested in the shape of an object) we must be clear what our precision 
estimates, expressed in terms of the dispersion matrix, are relative to. 
Thus in several close range applications the cosy notion of the photogrammetric 
precision being related to an external co-ordinate system founders. In such 
instances, the photogrammetrist is in a similar position to the geodesist 
and should, one could argue, apply conceptually similar solutions. What we 
require is a dispersion matrix which is free from variation in its elements 
caused by the arbitrary choice of restraining parameters (cf. the parameters 
we choose to fix to allow a relative orientation) and thus the concepts of 
internal precision (inner accuracy) and free network adjustment need to be 
appreciated. 
In a free network adjustment (where a conventional solution would be singular) 
the precision of a point's determination is related not to individual points, 
but to the network of points as a whole. This can be called "internal 
precision', a term which can also relate to non-singular systems. Conceptually, 
this can be likened to an absolute orientation in which every point acts 
as a control point. Methods relating to free network adjustments and 
internal precision have been well documented in the geodetic literature, with 
limited reference to the technique in papers on aerial photogrammetry by 
Meissl (1965), Ebner (1974) and Grun (1976). The application of such methods 
to close range photogrammetry is, however, more relevant. In executing the 
method a basic choice needs to be made between two approaches. In the first, 
the singular coefficient matrix of the normal equations N is augmented to 
incorporate a filter matrix G and a weight matrix P which eliminate the rank 
deficiency. The required dispersion matrix Q is then given by: 
-1 
N PG 
Q. m —- — 
GP O 
Alternatively, a dispersion matrix C is computed from the original system by 
imposing the minimum number of arbitrary constraints, and the required 
covariance matrix computed via the transformation: 
g=- men 
where 
mericpeiupockgt 
In either case the weighted trace of Q is a minimum. Further details of 
the method related to the multistation bundle adjustment can be found in 
Granshaw (1980). 
As an example of the application of the ideas of a free network adjustment 
and internal precision to the multistation bundle method, consider the 
simulated configuration illustrated in Fig. 1. Eight photographs are used 
to co-ordinate 96 targetted object points (16 regularly place on each face 
of the cube) of which targets 1 to 4 (Fig 1) are typical. Table 1 indicates 
the internal precision of this set up for three different cases. In case A 
no control information was used at all (ie the only measurements were photo 
co-ordinates to allow a multistation relative orientation), and so this is 
a true free network adjustment which could be used to evaluate the precise 
241