The article reports on a research project that investigates how the consistency of the 'manual indexing of images through lingustic means may be subject to less variability when carried out by domain experts rather than by lay viewers, in so far as experts more consistenlty select and name the physical characterisitcs of the image for indexing pruposes. The authors' domain of investigation is the 'collateral text' accompanying images of a crime scene conclude that the linguistic performance of the experts studied ban be characterised by a local grammar which is not applicable to the text produced by novices. The authors suggest, however, that the production of such collateral texts is responsive to training.