Title: Deep Learning Pipelines for Animal Re-Identification: A Survey from Detection to Population Inference
Authors: Serif Oyindamola Oyesiji, Toussida Fatah Tanguy Minoungou, Chukwudera Obumneke Anunagba, Harouna Wendpanga Yann Christian Sankara.
Volume: 10
Issue: 9
Pages: 322-334
Publication Date: 2026/09/28
Abstract:
Identifying individual animals from images is the quiet infrastructure of modern ecology: abundance estimation, movement analysis, social structure, and long-term monitoring all presuppose knowing which animal is which, and deep learning has transformed the identification from expert craft into a computational pipeline. A rapidly growing literature now spans that pipeline, yet its contributions are published stage by stage, detectors in vision venues, embedding losses in retrieval venues, species systems in ecology venues, and practitioners assembling working systems must synthesize across them. This survey reviews deep learning pipelines for animal re-identification by their anatomy, following the image from acquisition to identity. It first fixes the problem formulations, closed-set classification of known individuals, open-set identification with new-individual discovery, and verification, and the ecological constraints, imbalanced sightings, growing populations, and cross-deployment shift, that distinguish animal re-identification from its person and face antecedents. It then reviews each stage: detection and localization, where camera-trap-scale detectors and species-tuned localizers determine everything downstream; preprocessing, where quality filtering, viewpoint handling, and augmentation govern the effective training distribution; identity representation learning, the survey's core, organizing classification-based, metric-learning, and hybrid objectives, the treatment of patterned coats, fins, and faces, and the growing role of transfer and self-supervised pre-training under scarce labels; and matching, where ranking, gallery management, open-set thresholds, and scalable similarity search convert embeddings into decisions. Datasets and evaluation methodology are reviewed with attention to protocol pitfalls, identity leakage across splits foremost, and the deployment literature, platform systems and the coupling to capture-recapture inference, closes the pipeline. The survey consolidates challenges, open-set calibration, distribution shift, label scarcity, and long-tail sightings, and directions for the field.