Brazilian Road’s Animals (BRA): An Image Dataset of Most Commonly Run Over Animals

  • Gabriel S. Ferrante USP
  • Luis H. V. Nakamura IFSP
  • Fernando R. H. Andrade USP
  • Geraldo P. Rocha Filho UnB
  • Robson E. De Grande Brock University
  • Rodolfo I. Meneguette USP


In Brazil, about 475 million animals, including small and large species, are killed on the roads every year. Although this problem with wildlife can be considered recurrent, it involves significant investments and resources whose governments cannot afford to allocate to develop new mechanisms and technologies to identify, monitor, and guarantee the safety of animals in high-risk areas. In the field of computer vision, there are already datasets about animals in the literature, but not focusing on animals that suffer most accidents on highways. This work thus aims to present a new dataset of labelled images to address this gap, the BRA-Dataset. The labeling concerns five medium and large mammals that die the most on Brazilian roads, which can serve as a training dataset for future animal detection models on highways. The images were obtained using the proposed acquisition approach with a manual cleaning process; the dataset contain 1823 images in total and labelled in the YOLO Darknet and PASCAL VOC format. Also was effectuated experimentation on BRA-Dataset for validation of its performance on YOLO models training and tests. The results showed a high average precision achieved over the dataset in its first version without data augmentation or application process of improvement.
Palavras-chave: Training, Computer vision, Animals, Biological system modeling, Roads, Data models, Safety
Como Citar

Selecione um Formato
FERRANTE, Gabriel S.; NAKAMURA, Luis H. V.; ANDRADE, Fernando R. H.; ROCHA FILHO, Geraldo P.; GRANDE, Robson E. De; MENEGUETTE, Rodolfo I.. Brazilian Road’s Animals (BRA): An Image Dataset of Most Commonly Run Over Animals. In: CONFERENCE ON GRAPHICS, PATTERNS AND IMAGES (SIBGRAPI), 35. , 2022, Natal/RN. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2022 .