2.3 Image-matching software packages
We compared the performance of three feature-based image-matching software packages that differ in the underlying algorithms used to match individuals: I3S-Pattern (Reijns, 2014), WildID (Bolger et al., 2012) and Hotspotter (Crall et al., 2013). All three assist the user by listing potential matches for each image, ranked by a similarity score. The user then confirms which of these potential matches are true matches.
2.3.1 I3S-Pattern
I3S-Pattern uses the Speeded-up Robust Features (SURF) algorithm (Reijns, 2014; Bay et al., 2008) that selects key points and compares each image-pair in a dataset based on the size and position of these key points. The software requires the user to select three reference points per image, as well as the outline of the animal. As reference points, we used the base of the tail, the withers (i.e., the ridge between the shoulder blades), and the base of the neck (Figure S1).
2.3.2 Wild-ID
Wild-ID uses the Scale Invariant Feature Transform (SIFT) algorithm (Bolger et al., 2012; Lowe, 2004). It requires the user to input crops of the region of interest: the part of the animal which bears unique marks. The SIFT algorithm detects salient features regardless of their scale and viewpoint. For each image pair in a database, these features are compared, both in how the features look, and how the different features are positioned relative to each other. Based on these two characteristics, a goodness-of-fit score is computed per image pair.
2.3.3 Hotspotter
Hotspotter also uses the SIFT algorithm to conduct pairwise comparisons (Lowe, 2004; Crall et al., 2013). Users enter either entire pictures of individuals and select the rectangular region of interest, or image crops containing the region of interest. Hotspotter supplements the pairwise comparisons used by Wild-ID with a one-vs-many approach that uses a Local Naive Bayes Nearest Neighbour method (McCann & Lowe, 2012) to take all of the images in the database into account when computing similarity scores.