Benchmarking wild bird detection in complex forest scenes

Camera traps are widely used for wildlife monitoring and making informed conservation and land-management decisions, but the resulting 'big data' are laborious to process. Deep learning-based methods have been adopted for wildlife detection in camera traps. However, these methods detect large mammals in uncomplicated scenes, where powerful deep-learning models work effectively. Few studies have been conducted to develop artificial intelligence for recognizing wild birds that live in complicated field scenes with protective colors and small sizes. Here we used a dataset of 9717 images from 15 bird species based on camera traps to test 8 object detection algorithms (Faster RCNN, Cascade RCNN, RetinaNet, FCOS, RepPoints, ATSS, Deformable-DETR, and Sparse RCNN) and assess their performance. We also explored the effect of different backbones on model accuracy. Among them, the Cascade RCNN model performs best, with a mAP of 0.693 in model capabilities. Models perform differently in certain species, and backbones significantly affect the accuracy of the model. Cascade RCNN utilizing the Swin-T backbone is the best-performing combination, with a mAP of 0.704. This study could help researchers identify birds efficiently and inspires research on wildlife recognition in complex ecological settings.