BRAGAN: a GAN-augmented dataset of Brazilian roadkill animals for object detection

BRAGAN is a new dataset of Brazilian wildlife developed for object detection tasks, combining real images with synthetic samples generated by Generative Adversarial Networks (GANs). It focuses on five medium and large-sized mammal species frequently involved in roadkill incidents on Brazilian highways: lowland tapir (Tapirus terrestris), jaguarundi (Herpailurus yagouaroundi), maned wolf (Chrysocyon brachyurus), puma (Puma concolor), and giant anteater (Myrmecophaga tridactyla). Its primary goal is to provide a standardized and expanded resource for biodiversity conservation research, wildlife monitoring technologies, and computer vision applications, with an emphasis on automated wildlife detection. The dataset builds upon the original BRA-Dataset by Ferrante et al. (2022), which was constructed from structured internet searches and manually curated with bounding box annotations. However, while the BRA-Dataset faced limitations in size and variability, BRAGAN introduces a new stage of dataset expansion through GAN-based synthetic image generation, substantially improving both the quantity and diversity of samples. In its final version, BRAGAN comprises approximately 9,238 images, divided into three main groups: Real images — original photographs from the BRA-Dataset. Total: 1,823. Classically augmented images — transformations applied to real samples, including rotations (RT), horizontal flips (HF), vertical flips (VF), and horizontal (HS) and vertical shifts (VS). Total: 7,300. GAN-generated images — synthetic samples created using WGAN-GP models trained separately for each species on preprocessed subsets of the original data. All generated images underwent visual inspection to ensure morphological fidelity and proper framing before inclusion. Total: 115. The dataset follows an organized directory structure with images/ and labels/ folders, each divided into train/ and val/ subsets, following an 80–20 split. Images are provided in .jpg format, while annotations follow the YOLO standard in .txt files (class_id x_center y_center width height, with normalized coordinates). The file naming convention explicitly encodes the species and the augmentation type for reproducibility. Designed to be compatible with multiple object detection architectures, BRAGAN has been evaluated on YOLOv5, YOLOv8, and YOLOv11 (variants n, s, and m), enabling the assessment of dataset expansion across different computational settings and performance requirements. By combining real data, classical augmentations, and high-quality synthetic samples, the BRAGAN provides a valuable resource for wildlife detection, environmental monitoring, and conservation research, especially in contexts where image availability for rare or threatened species is limited.BRAGAN是一款专为目标检测任务开发的巴西野生动物全新数据集，融合了真实图像与由生成式对抗网络（Generative Adversarial Networks，GANs）生成的合成样本。该数据集聚焦常卷入巴西高速公路路杀事件的5种中大型哺乳动物物种：低地貘（Tapirus terrestris）、细腰猫（Herpailurus yagouaroundi）、鬃狼（Chrysocyon brachyurus）、美洲狮（Puma concolor）以及大食蚁兽（Myrmecophaga tridactyla）。其核心目标是为生物多样性保护研究、野生动物监测技术与计算机视觉应用提供标准化且经扩容的资源，重点面向自动化野生动物检测任务。 该数据集基于Ferrante等人2022年提出的原始BRA-Dataset构建，后者通过结构化网络搜索获取数据，并经人工审核并标注边界框。但原始BRA-Dataset存在样本数量与多样性不足的局限，BRAGAN则通过基于GANs的合成图像生成技术实现了数据集的进一步扩容，大幅提升了样本的数量与多样性。最终版本的BRAGAN共包含约9238张图像，分为三大主要类别： 1. 真实图像——源自BRA-Dataset的原始照片，共计1823张。 2. 经典增强图像——对真实样本应用各类变换操作，包括旋转（RT）、水平翻转（HF）、垂直翻转（VF）、水平平移（HS）与垂直平移（VS），共计7300张。 3. GAN生成图像——通过针对每个物种分别在原始数据预处理子集上训练的沃瑟斯坦GAN带梯度惩罚（Wasserstein GAN with Gradient Penalty，WGAN-GP）模型生成的合成样本。所有生成图像在纳入数据集前均经过人工视觉检查，以确保形态保真度与合理构图，共计115张。 该数据集采用标准化目录结构，包含images/与labels/两个文件夹，二者均按照80-20的比例划分为训练集（train/）与验证集（val/）子集。图像采用.jpg格式存储，标注文件遵循YOLO（You Only Look Once）标注规范，以.txt格式存储，格式为：class_id x_center y_center width height，坐标均已归一化。文件命名规则明确编码了物种与增强类型信息，以保障实验可复现性。 BRAGAN兼容多款目标检测架构，已在YOLOv5、YOLOv8及YOLOv11（包含n、s、m三个变体版本）上完成评估测试，可用于评估不同计算配置与性能需求下的数据集扩容效果。 通过融合真实数据、经典增强样本与高质量合成样本，BRAGAN为野生动物检测、环境监测与保护研究提供了极具价值的资源，尤其适用于珍稀或濒危物种图像样本获取受限的场景。