Long-Tailed Metrics and Object Detection in Camera Trap Datasets

With their advantages in wildlife surveys and biodiversity monitoring, camera traps are widely used, have been used to gather massive amounts of animal images videos. The application deep learning techniques has greatly promoted the analysis utilization trap data management conservation. However, long-tailed distribution dataset can degrade performance. In this study, for first time, we quantified long-tailedness class object/box-level scale imbalance datasets. dataset, problem is prevalent severe, terms scale. worse imbalance, too few samples small objects, making more challenging. Furthermore, BatchFormer module exploit sample relationships, improved performance general object detection model, DINO, by up 2.9% 3.3% imbalance. experimental results showed that relationship was simple effective, improving but it could not make low number objects dataset.