Applying computer-aided photo-identification to messy datasets: a case study of Thornicroft's giraffe (Giraffa camelopardalis thornicrofti)

Digital photography enables researchers to rapidly compile large quantities of data from individually identifiable animals, and computer software improves the management of such large datasets while aiding the identification process. Wild-ID software has performed well with uniform datasets controlling for angle and portion of the animal photographed; however, few datasets are collected under such controlled conditions. We examined the effectiveness of Wild-ID in identifying individual Thornicroft's giraffe from a dataset of photographs (n=552) collected opportunistically in the Luangwa Valley, Zambia from March to October 2009. We assessed the programme's accuracy in correctly identifying individuals and the effect of five image quality factors on identification success: blurriness, background type and complexity, amount of sky and the presence of other giraffe. The programme correctly identified individuals in 71.6% of photographs. Background complexity was the only significant variable affecting identification success and removing background imagery reduced identification error by 52.8% (from 28.4 to 13.4%). Our results indicate higher levels of error than previously reported for Wild-ID. However, they also suggest the programme is an effective tool for quickly identifying individuals in large field datasets, especially if photograph backgrounds are removed beforehand and postanalysis visual verification is performed.