Résumé
Preserving endangered species is a critical component of maintaining a balanced and healthy ecosystem. Animal pose, especially for rare animals, allows an understanding of various aspects of biology and ecology, including but not limited to individual animal behavior analysis and study of migration patterns. Using the small-scale dataset from (i.e., red-list species) monitoring efforts of Eurasian lynx (Lynx lynx), we provide a comprehensive guide to a simple yet effective 2D pose estimation suitable for endangered species. We showcase the contribution of a variety of methods and their influence on the performance in terms of AP, AP 0.75 , AP 0.85 , and PCK 0.05 . Our experiments provide a hitchhiker's guide to (i) pre-trained model selection, (ii) model pre-training and fine-tuning, (ii) augmentation strategies, (iii) training hyper-parameters settings, (iv) number of required real data, and (v) use of synthetic data. Using all the bells and whistles and HRNet-w32, we achieved 0.855AP and 0.936PCK 0.05 lowering the relative error of a pretrained model by more than 50%. Last but not least, we have developed a system for photorealistic synthetic camera trap data generation. The system is available at: https://github.com/strakaj/Synthetic-animal-pose-generation.git.