Benchmarking NACTI Species Recognition in Long-Tailed Regimes

This paper addresses the persistent challenge of long-tailed class imbalance in wildlife monitoring, specifically within the North American Camera Trap Images (NACTI) dataset. With over half of the 3.7 million images belonging to a single species (the domestic cow), standard machine learning models often become biased toward majority classes, failing to identify rare or endangered species. The authors systematically evaluate various Long-Tail Recognition (LTR) methodologies, including Focal loss, Weighted Cross-Entropy (WCE), and Label-Distribution-Aware Margin (LDAM) loss, integrated with advanced learning rate scheduling and optimization techniques.

The study demonstrates that integrating a learning rate scheduler (specifically ReduceLROnPlateau) with LTR-sensitive loss functions provides substantial performance gains. The authors' optimized configuration, which pairs LDAM loss with an AdamW optimizer and a learning rate scheduler, achieved a state-of-the-art 99.40% Top-1 accuracy on the NACTI test split. Furthermore, the model showed improved generalization capabilities across three independent, reduced-bias out-of-distribution (OOD) test sets compared to standard cross-entropy approaches.

Accurate species classification is critical for biodiversity conservation and ecological monitoring. By providing a reproducible, high-performance pipeline and open-sourcing the code, weights, and dataset splits, this work offers a robust framework for researchers to improve minority-class detection. However, the authors emphasize that while optimization techniques can boost performance, they do not resolve the fundamental representational bottlenecks that lead to failure when models encounter severe domain shifts, such as night-time captures or motion blur, for rare species.