discussion / Camera Traps  / 29 March 2017

Data Needed: Camera trap images for machine vision training

Hi everyone,

I'm working on an object detection system for detecting elephants! It's detailed here https://www.wildlabs.net/community/thread/284 and at my project page here too https://hackaday.io/project/20448-elephant-ai . At present I am still attempting to train an object detector using Histogram of Oriented Gradients (HOG) and Linear Support Vector Machines (SVM) implemented using python w open-source libraries. Although I might move to TensorFlow.

So, for training my object detector, I need many images of both elephants and other 'not-elephant' large mammals (specifically: macaques, langurs, gibbons, wolves, jackals, foxes, sun bears, black and brown bears, sloth bears, otters, civetss, mongooes, tigers, leopards, other smaller felines, rhinos, hogs, boars, deer, yaks, buffalo, antelopes). It's especially hard differentiating rhinos from elephants. That is a big problem! 

It would be best if animals were at range of 2m+ from camera. At the moment I would like to do a training run with 5000 non-elephants and 1000 elephants! Just from IR-filtered (daylight) cameras at the moment. 

If anyone could help out by sharing their images: either by uploading images via ftp or sharing/uploading via dropbox, etc. that would be really great! 

Thanks for reading!

Hi Peter, 

I manage the Snapshot Serengeti Project -- we have over 31,000 images that contain elephants, and 100,000s of images that contain other species. If you are interested in using these images, please email me at palme516 at umn dot edu, we can set up an FTP or Globus transfer! 


Have you looked at the ImageNet dataset? I'm not sure if you can download directly from ImageNet anymore, but you can get links to a lot of Internet images. They show an elephant data set of almost 1400 and a rhino dataset of nearly 1500.

Perhaps you could also consider using a Convolutional Neural Network which was already trained on the imagenet dataset. You may find one that already has a classifier for elephants. For example, the 2015 challenge included elephants. You can probaly find a pretrained GooLeNet or ResNet that can detect elephants.

Best of luck!



We would be keen to see how you go, we are developing similar systems for a range of species and hope to submit a couple of publications very soon.


Are you running this stuff on OpenCV/Raspberry Pi?

I wa contemplating whether such a combination might be able to detect bipeds (humans) vs quadrupeds... and of course, how much power it would require.