Detecting Plains and Grevy's Zebras in the Realworld
Abstract
Photographic censusing can be partly automated by leveraging the power of computer vision detection algorithms. Detecting zebras in the real world can be challenging due to varying viewpoints of the animal, natural and artificial occlusions, and overlapping animals. To address these challenges, we evaluate three detection algorithms: Hough Forests by [8], the YOLO network by [20], and Faster R-CNN [21]. We train the detectors on a soon-to-be-released dataset of 2,500 images containing 3,541 bounding boxes of plains zebras (Equus quagga) and 2,672 bounding boxes of Grevy's zebras (Equus grevyi). The detection errors are analyzed by species, viewpoint, and density (the number of bounding boxes per image). The best detector in our evaluation reports a detection mAP of 55.6% for plains and 56.6% for Grevy's.
Cite
Text
Parham and Stewart. "Detecting Plains and Grevy's Zebras in the Realworld." IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, 2016. doi:10.1109/WACVW.2016.7470122Markdown
[Parham and Stewart. "Detecting Plains and Grevy's Zebras in the Realworld." IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, 2016.](https://mlanthology.org/wacvw/2016/parham2016wacvw-detecting/) doi:10.1109/WACVW.2016.7470122BibTeX
@inproceedings{parham2016wacvw-detecting,
title = {{Detecting Plains and Grevy's Zebras in the Realworld}},
author = {Parham, Jason R. and Stewart, Charles V.},
booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision Workshops},
year = {2016},
pages = {1-9},
doi = {10.1109/WACVW.2016.7470122},
url = {https://mlanthology.org/wacvw/2016/parham2016wacvw-detecting/}
}