Wild Animal Facing Extinction
Description
Wild animal poaching is quite severe, especially in Africa, where South Africa occupies 82% of the world’s rhino community. Africa is poised to be leading in wildlife and poaching badly influences the continent tourism business, as it adds to the extinction of wild animals. Governments and numerous Non-Government Organizations (NGOs) are spending a lot of money and time protecting wild animals from poaching through various methods, such as upgrading fencing systems and employing video surveillance and monitoring systems. The deep-learning-based Computer Vision (CV) solutions depend severely on the large volume of annotated image data. In computer vision-based conservation studies, having a wild animal dataset is an additional advantage for making informed dynamic decisions. Traditionally, most datasets are built through manual annotation, where labelImg is used to draw boundary boxes. However, manual image annotation is very time-consuming, particularly when images are countless and that ends up becoming an expensive task because such assignments will need to be outsourced. We investigate a framework for semi-auto annotation based on boundary box labels but not tool development. We contribute by manually annotating a small dataset and training a model and semiauto annotating a large scale dataset using the trained model and later correct the miss-classified objects. Furthermore, we contribute a framework that minimizes annotation time and can be used in any dataset construction. Our wild animal dataset is a contribution to conservationists and wild animal literature, as wild animal research faces limitations in computer vision. In initializing the process, the images are collected through a search of words from both Google and Baidu search engines manually for six classes; during the manual collection, we carefully do that through web-scraping, which helps in avoiding wrongly categorized images and avoid saving similar images. A small set of images is then randomly chosen for annotation and trained through YOLOv5 to produce a customized model whose weights will suggest new boundary boxes for the remaining large set of images that were not annotated when both weights and a large set of images are fed to labelGo, to complete the process automatically. Human gets involved by correcting the boundary boxes that have class errors in correcting the misclassification of objects, and after the corrected images are saved to the folder that has images that were labeled by labelGo, in complementing semi-auto annotation subsequently the approach minimizes the time-consuming during manual annotation. Wild animals Number of images Rhino 1434 Cheetah 1183 Elephant 1218 Lion 1305 Zebra 1377 Giraffe 1232 Set Number of Images Train 6505 Test 621 Validation 623 Total 7749 Image = .jpg & Annotation = .txt