GCN-ID: A Benchmark Dataset for Great Crested Newt Re-Identification Using AI Foundation Models

Great crested newts (Triturus cristatus) are a long-lived species of Class Amphibia and are protected by wildlife legislation across Europe. A key bioindicator of the health of the environment, great crested newts (GCN) have experienced significant population declines. Individual GCNs can be differentiated based on their unique orange-yellow underbelly markings with black blotching patternisation, but manually identification is time-consuming and requires domain knowledge. We collect and publicly release a comprehensive dataset of images (n = 1232) and videos (n = 1233) of GCNs(n=206different individuals) that can be used to develop AI models to aid in the autonomous identification of individuals of this species. Our data set comes with bounding boxes, segmentation masks, and multiple query-database splitting strategies. In addition, we propose two novel splitting methods that can be used to estimate performance inflation in the context of animal re-identification by: i) exploiting different biases in re-identification models; and ii) retrieving outlier samples with maximal difference in capture conditions. Using such splits, we show that the performance of State-of-the-Art (SoTA) foundation models for animal re-identification drops significantly (up to 36 percentage points for top-1 accuracy) when compared with random test splits. Nonetheless, we obtain a top-1 accuracy of 63% (top-5 of 83%) for the MegaDescriptor and a top-1 accuracy of 73% (top-5 of 93%) for the MiewID using the hardest bias-exploiting split with minimal data processing. This benchmarking exercise offers awareness of the effects of biases on accuracy estimation, but also encouraging baselines for future research on GCN re-identification based on AI foundation models.