I did a quick evaluation of whether Google Cloud Vision can be used off the shelf to automatically identify animals from camera trap photos. I wrote a blog post about it here.
Summary: If the animal photo is clear, then you could expect 80% to be identified; if photos are not clear, this drops all the way down to 11%.
12 April 2016 2:41pm
Interesting. i remember a paper came out highlighting a similar system. I feel like within a year or two automatic animal identification for camera traps will be come a reality.
20 April 2016 11:34am
Great article, thanks for sharing Aditya! Given the (mildly alarming) amount of time it took me to actually spot the leopard in your 'very difficult' example, I'm actually rather impressed that Google Cloud Vision was able to identify the presence of an animal at all in photos falling under that classification.
7 August 2016 11:25pm
An update to the automated species identification debate:
A paper has recently come out which used deep learning ("very deep convolutional networks") and managed 89% accuracy for the Snapshot Serengeti Zooniverse dataset, IF the image was first manually cropped around the animal. Seriously, who has time to do that? If the image remained uncropped they managed a woeful 35% accuracy.
Perhaps we have a long wait ahead of us for this to become a practical reality?
John Probert
World Wide Fund for Nature/ World Wildlife Fund (WWF)