Long-tailed Species Recognition in the NACTI Wildlife Dataset

TL;DR

The paper tackles the challenge of long-tail recognition in wildlife image data from the NACTI dataset, where a single head class dominates the distribution. It evaluates a range of LTR losses (focal, WCE, LDAM) with learning-rate scheduling in a scalable multi-GPU ResNet-50 pipeline, comparing against a standard cross-entropy baseline. The best configuration—LDAM loss with AdamW and ReduceLROnPlateau scheduler—achieves 99.40% Top-1 accuracy on the biased NACTI test set and 99.22% on tail classes, while generalisation under distribution shift is tested on a reduced-bias ENA-derived set where accuracy reaches 52.55%. The study highlights strong gains from LTR scheduling and LDAM in the wildlife domain, but also reveals persistent tail-class failures under domain shift and stresses the need for robust OOD evaluation and data-centric remedies. The authors provide fully reproducible resources, including dataset splits, code, and network weights, to foster ongoing advancements in camera-trap species recognition.

Abstract

As most ''in the wild'' data collections of the natural world, the North America Camera Trap Images (NACTI) dataset shows severe long-tailed class imbalance, noting that the largest 'Head' class alone covers >50% of the 3.7M images in the corpus. Building on the PyTorch Wildlife model, we present a systematic study of Long-Tail Recognition methodologies for species recognition on the NACTI dataset covering experiments on various LTR loss functions plus LTR-sensitive regularisation. Our best configuration achieves 99.40% Top-1 accuracy on our NACTI test data split, substantially improving over a 95.51% baseline using standard cross-entropy with Adam. This also improves on previously reported top performance in MLWIC2 at 96.8% albeit using partly unpublished (potentially different) partitioning, optimiser, and evaluation protocols. To evaluate domain shifts (e.g. night-time captures, occlusion, motion-blur) towards other datasets we construct a Reduced-Bias Test set from the ENA-Detection dataset where our experimentally optimised long-tail enhanced model achieves leading 52.55% accuracy (up from 51.20% with WCE loss), demonstrating stronger generalisation capabilities under distribution shift. We document the consistent improvements of LTR-enhancing scheduler choices in this NACTI wildlife domain, particularly when in tandem with state-of-the-art LTR losses. We finally discuss qualitative and quantitative shortcomings that LTR methods cannot sufficiently address, including catastrophic breakdown for 'Tail' classes under severe domain shift. For maximum reproducibility we publish all dataset splits, key code, and full network weights.

Figures (6)

• Figure 1: Representative Samples from the NACTI Dataset. Shown is one representative image per species class from the overall 48 classes in total, excluding empty and vehicle frames. Note challenging occlusions, background clutter, motion blur, infrared shots, and small animal areas. Red bounding boxes indicate detections using MegaDetectorV6, applied during the pre-processing stage. The NACTI dataset is long-tailed with over half of the samples belong to a single species (domestic cow). Following (Perrett et al.,2023), 5 classes form the head, 37 form the ‘long tail’, and 6 few-shot classes sit at the end of the species distribution.
• Figure 2: LTR Performance Improvements across NACTI Species vs. Baseline. Per-class Top-1 accuracies on the NACTI dataset comparing the baseline model (cross-entropy with Adam) against our best-performing configuration, which combines a Label-distribution-aware margin (LDAM) loss, an alternative AdamW optimiser, and a learning-rate scheduler. The results are reported on a 10% random test split. The blue bars indicate classes with a positive improvement relative to the baseline.
• Figure 3: NACTI Species Distribution after Sample Balancing. Logarithmic visualisation of the cardinality of all 48 (sub)species classes of NACTI after capping classes at 100k. Colouration of the class distribution after perrett2023useheadimprovinglongtail breaking the distribution into 'Head' (blue), 'Tail' (pink), and 'Few-shot' species.
• Figure 4: Training Dynamics. Column (a) reports accuracy whilst column (b) reports on loss on one of the training runs with scheduling. Within each column, the left plot shows train and validation, right plots show the train-validation gap. Note the dashed lines marking early-stopping point.
• Figure 5: Qualitative Example regarding Domain Shift for 'Horse'. Train-test contrast of some samples from the horse class where we observed catastrophic performance breakdown. Training images are often blurry or occluded, whereas reduced-bias test images involve night scenes, partial views and unusual poses.