Descriptor: A Brazilian Roadkill Animals for Object Detection GAN-Augmented Dataset (BRAGAN)

Collisions between vehicles and wild animals pose a major threat to biodiversity and road safety in Brazil, resulting in the death of millions of animals each year. Despite the seriousness of this issue, there are still a few open datasets focused on computer vision research on collision detection. To fill this gap, we present BRAGAN, a new GAN-augmented dataset derived from the Brazilian Road Animals (BRA-Dataset). Although the BRA-Dataset provided an initial collection of real images of five species of medium and large mammals most frequently involved in roadkill, its limited size highlighted the difficulty of acquiring such images in natural conditions. BRAGAN expands this resource by leveraging several GAN architectures, trained separately for each species on carefully preprocessed subsets of data. The resulting dataset integrates real samples and those generated by the best network according to qualitative criteria, with Wasserstein Gradient Penalty (WGAN-GP) standing out, increasing the variability in the animals’ appearance, posture, and background consistency. BRAGAN was designed to train and compare object detection models, such as YOLO, and is publicly available to support applications in computer vision, wildlife monitoring, conservation technology, and road safety planning.