3.1 Automatic detection and aspect-ratio filtering
The Microsoft AI for Earth MegaDetector, which is designed to
automatically detect and crop animals from images, produced 2652 crops
from the test dataset of 1306 images, meaning that on average,
approximately two detections were made per image. Of these, 2523 crops
contained a wild dog (true positive rate = 0.951), 129 were false
detections, such as rocks or vegetation (false positive rate = 0.049),
while 531 wild dogs were not successfully detected (false negative rate
= 0.174). However, only five of these false negatives were images
suitable for image-matching. By contrast, the flank of the wild dog was
not visible in the other 526 false negatives. In total, 722 crops were
suitable for identification using image matching software, of which five
were not detected by the automated processing (false negative rate =
0.007). For the 2652 crops that were produced by the MegaDetector, all
crops considered suitable for identification on visual inspection had an
aspect-ratio between 0.65 and 2.25. Applying the MegaDetector to the
entire image dataset (n = 11205), as opposed to the test dataset,
resulted in 21745 crops. Of these, 5788 (21%) fell outside the suitable
range of aspect ratios and were therefore removed from the dataset.