Harvey et a! : Comparison of estimates of reef-fish lengths made by divers and a stereo-video system 
65 
silhouette to the stereo-video rig was greater than 50 de- 
grees, resulting in a deterioration of the accuracy and pre- 
cision of measurements (Harvey and Shortis, 1996). Thus, 
76 pairs of images were retained for analysis, from each of 
which ten measurements were made. 
To test whether measurements made by an inexperi- 
enced volunteer using the stereo-video software would dif- 
fer from those made by an experienced operator, two vol- 
unteers were given brief instructions on how to operate 
the stereo-video software and were asked to make mea- 
surements of the silhouettes. Measurements made by the 
inexperienced operators were then compared with those 
made from the same images by an experienced operator 
To determine whether changes in the water clarity af- 
fect the accuracy and precision of length estimates made 
by the stereo-video system, a further three transects were 
sampled in a saltwater swimming pool at the PML during 
August 1994. Water clarity was measured with a tape mea- 
sure to record the distance to the farthest, clearly visible 
silhouette. A silhouette was clearly visible at up to 25 me- 
ters in the swimming pool at the University of Melbourne 
and up to 5.5 meters in the saltwater pool at PML. 
We expected that the accuracy of measurements ob- 
tained with the stereo-video system would improve with 
operator experience. To evaluate the effect of experience, 
the same operator remeasured the images recorded in the 
swimming pool in Melbourne, approximately one year af- 
ter the initial analysis. In the intervening period the op- 
erator routinely used the stereo-video system for analysis 
of fish lengths and was by far the most experienced person 
using the system. 
Analysis of data 
Four measures of error were used to summarize the accu- 
racy of each length estimate. If the observed fish length is 
O and the true length is T, these four measures are 
Error: E = O - T\ 
Relative error: RE = (O - T)/T = E!T\ 
Absolute error: AE = j O - T | = | E | ; 
Relative 
absolute error: RAE = \ 0 - T\/T = \E \ IT = \ RE \ . 
The four measures provide different types of information. 
The error E will be positive or negative according to 
whether the observed length is an overestimate or an 
underestimate. If the mean of E is close to zero, the 
reason might be that all the estimates are accurate or that 
some are overestimates and others are underestimates to 
approximately the same degree. The absolute error AE 
ignores the direction of the error, and thus would provide 
different mean values for these two scenarios. The rela- 
tive errors RE and RAE are of interest because it might be 
expected that these would be consistent across a range of 
fish lengths. The measure of accuracy used by St John et 
al. (1990) was simply RE + 1. 
The results for the novice and experienced scientific div- 
ers were summarized prior to further analysis by calcu- 
lating, for each silhouette and each observer, the mean 
length estimate over all transects. The repeat transects 
were therefore used solely to provide a reliable estimate of 
the error made by each observer on each silhouette. If the 
transects were truly independent, use of the mean would 
tend to provide a conservative estimate of the error a diver 
would make on one transect. Because the transects were 
swum in quick succession, they were not independent and 
therefore the degree to which their error was underesti- 
mated should be small. Each mean length estimate was 
then used to derive the four summaries of error, for each 
silhouette and each observer. For each measure of error, a 
two-factor analysis of variance was performed, the factors 
being type of observer and silhouette length. The individual 
observers provided the replication needed for this analysis. 
To summarize the estimates made by using the stereo- 
video, the mean length estimate over all ten measure- 
ments was calculated (for each silhouette, transect, and 
operator). The mean of these over all five transects was 
then calculated for each silhouette and operator and con- 
verted to the four summary measures of error. For each 
measure of error, a one-factor analysis of variance was per- 
formed, with silhouette length being the factor. The indi- 
vidual operators (one experienced and two inexperienced) 
provided the replication needed for this analysis. 
The two types of diver data and all video data were ana- 
lyzed separately, rather than in a single analysis of vari- 
ance because for each silhouette the variation between 
video operators was found to be much lower than that 
between experienced scientific divers and novice divers, 
making the usual assumption of equal variance invalid 
(Underwood, 1981). Comparison of the video technique 
with that of the experienced scientific divers and novice 
divers was therefore made by comparing 95% confidence 
intervals. Tbe purpose of the analysis of variance for the 
video data was to provide an estimate of the standard er- 
ror associated with the mean (for each measure of error) 
over all silhouettes. 
To compare the data of experienced and inexperienced 
stereo-video operators, the mean estimated length over the 
ten repeat measurements obtained per image was calcu- 
lated. These means were then converted to the four mea- 
sures of error. Each measure of error was then analyzed by 
using a two-factor analysis of variance, with operator and 
silhouette as the factors. In this analysis, therefore, unlike 
those for comparing estimates from divers with those pro- 
duced with the stereo-video system, the transects provided 
replication against which to compare operators. This anal- 
ysis is reasonable, because in comparing the operators, the 
transects can be regarded as independent. 
To compare the use of a stereo-video system under the 
two water-clarity conditions, a three-factor analysis of 
variance was performed on the individual errors. In this 
analysis, the factors were water clarity, silhouette, and 
transect (nested within each combination of water clarity 
and silhouette). Prior to this analysis, the estimates for 
one of the silhouettes in 5.5-m water clarity were removed 
because this silhouette was 6.6 meters from the transect 
and could not be seen clearly. 
For the 1994 and 1995 stereo-video data used to assess 
operator experience, a three-factor analysis of variance 
