Promising results have been achieved so far, though further development is needed. The first result in Fig. 2 illustrates a successful identification; the third is less satisfactory. The model in § 4 describes a subset of image features given the data. It would be preferable to have a more complete probabilistic description, with a more careful physical analysis of the lighting effects taking into account the 3D nature of fish (see [ 10 ], for example).
It is observed in the data that various object features (both biological and visual) are sometimes absent in individual cases. For example, fins are effectively absent when flattened against the fish body, and eyes can be identified only in a subset of the images. It may be beneficial to model the presence or absence of such features explicitly.
Although our dataset was limited to still images, newer data involving image sequences raises the possibility of a more powerful analysis incorporating tracking (see [ 15 ], for example). Also, the technique must be extended to deal with images containing multiple objects. MCMC is well-suited to such questions: for example, see [ 6 ] and [ 7 ] for MCMC methods dealing with multiple objects, and [ 12 ] for the treatment of occlusion of overlapping objects.
K De Souza