/BRA-Dataset

Project for Master's Degree in Computer Science and Matemathical Computational

Primary LanguagePython

BRA-Dataset

BRA-Dataset is a dataset about the animals that most suffer accidents on Brazilian highways. It has 5 classes of medium and large animals, with 1823 imagens. It is free and open-source.

NEWS

Data Augmentation script with geometric transformations

Tutorial for COCO format conversion

Metrics achieved:

Technical features:

1823 Images Total

2060 Labels Total

Classes:

Tapir----------------------------------------------> 403 Images / 478 Labels

Jaguarundi-----------------------------------------> 311 Images / 341 Labels

Maned Wolf-----------------------------------------> 340 Images / 410 Labels

Puma-----------------------------------------------> 476 Images / 495 Labels

Giant Anteater-------------------------------------> 293 Images / 336 Labels

Convert darknet format for COCO format

clone the repository YOLO2COCO

With the diretory struture

Dataset ---images ---labels ---class.names ---gen_config.data ---train.txt ---valid.txt

execute the command in YOLO2COCO folder: python darknet2coco.py --data_path ../Dataset/gen_config.data

Videos for test your detector

https://mega.nz/file/3xREFRrJ#pAmxuOanRehFlo75LUAfBokq2huZ22jUfcuPRFNQ5ow - For Tapir Class with oclusion problem

https://mega.nz/file/PhhFhSgQ#1NScqSqNIOcpWgE7JR_ntbogw-b7ecKdle0mn6PRSjI - For Giant Anteater with small object and camouflage

About

This project is a master's project in computer science and computational mathematics at the University of São Paulo - Brazil (PPG-CCMC, ICMC-USP)

Authors: Gabriel Souto Ferrante

Paper: Brazilian Road Animals (BRA): An Image Dataset of Most Commonly Run Over Animals | DOI: 10.1109/SIBGRAPI55357.2022.9991774