Animal Re-Identification

Animal re-identification asks whether observations made at different times show the same animal, supporting wildlife monitoring, ecological studies, and livestock management without relying only on tags. Common systems learn identity embeddings with convolutional neural networks or vision transformers and compare coat patterns, faces, muzzles, or other markings; detection, segmentation, local-patch matching, and graph clustering help with crowded scenes and open-set galleries. Other approaches use temporal tracking, location or environmental metadata, audio, and vision-language features, while synthetic views and degradation-aware training target changes in viewpoint, image quality, species, and time. Evaluation spans species-specific retrieval, verification, and tracking benchmarks, but transfer across cameras and individuals remains difficult, and high scores do not guarantee calibrated decisions or ecologically useful error rates.