PACTraD: Multi-Task Benchmark and Domain-Specific Attention Mechanism for Wildlife Classification, Detection, and Segmentation in Protected Area Camera Trap Data

Camera traps are a key tool for non-invasive wildlife monitoring, yet models trained on existing benchmarks often generalize poorly to new regions due to domain shift, class imbalance, and natural variability in illumination, season, and background. We introduce PACTraD, a benchmark of 116,973 annotated camera trap images collected in a protected area of the Baikal region across eight classes (deer, no_animals, bear, boar, hare, badger, lynx, wolf), with a pronounced long-tailed distribution (deer: 73.1%; lynx and wolf: 0.3% and 0.2%) and substantial seasonal and lighting variability, including infrared night imagery. The benchmark supports three tasks under a unified, reproducible protocol: classification, object detection, and semantic segmentation. We further propose a domain-specific attention mechanism augmenting self-attention with eleven learnable priors capturing contour, foreground/background separation, bilateral symmetry, discriminative part regions, texture frequency, confidence weighting, motion consistency, gradient enhancement, scale, color invariance, and shape compactness. Evaluated on six backbones for classification and detection and five for segmentation, against no-attention baselines and three generic attention modules (SE, SA, CBAM), it consistently performs best, raising weighted F1 for classification from 0.943 to 0.964 (DINOv3), AP50 for detection from 0.876 to 0.903 (InternImage-H), and mIoU for segmentation from 0.805 to 0.834 (SegMAN), with component-level ablations revealing task-specific contributions of each prior. However, performance on rare species remains limited (macro F1: 0.829 for classification), indicating that further research is needed to address long-tail recognition in camera trap data.