Key points are not available for this paper at this time.
Biodiversity conservation hinges on the ability to monitor endangered wildlife within forest ecosystems. Traditional survey methods are often impractical due to the elusive nature of these species and the difficulty of reaching their habitats. Unmanned Aerial Vehicles (UAVs) have emerged as a powerful and non-invasive alternative. It remains a technical bottleneck that current automated detection models struggle with complex forest canopies. High-quality data are essential for effective, robust modeling. Most existing UAV datasets focus on charismatic mammals in open savannas, leaving a critical gap for species inhabiting complex forest canopies. To address this disparity, we present the Aerial Endangered Wildlife Dataset (AEWD), curated for critical monitoring in natural environments with high clutter. It includes 7483 high-resolution images, with 27,194 annotated instances of four key species: the Amur Tiger ( Panthera tigris altaica ), Giant Panda ( Ailuropoda melanoleuca ), Golden Snub-nosed Monkey ( Rhinopithecus roxellana ), and Sichuan Takin ( Budorcas tibetanus ). Unlike conventional benchmarks, AEWD assesses detection difficulty by considering five detailed attributes: bounding box size, instance scale, target density, vegetation coverage, and occlusion degree. These metrics reflect real-world challenges, such as indistinct features and low contrast. By evaluating twelve mainstream detection models, we establish a performance baseline to catalyze future research in UAV-based wildlife conservation. • AEWD provides a specialized benchmark for rare endangered species detection. • Acquired 7483 UAV images across remote and inaccessible forest terrains. • Over 27,000 instances are expert-labeled in natural habitats. • Fine-grained attributes quantify weakly observable UAV challenges.
Ma et al. (Mon,) studied this question.