Thanks Ben. I'll see what I can do.

Hi Michelle,

I had a group of undergrads help me with a 40,000-image dataset a few years back. We used the TEAM network Wild.ID program, so each photo that was tagged indicated who tagged it. That was helpful for checking quality later on. For our common, unmistakeable species (e.g. whitetail deer), I didn't require a second identification, but for more challenging groups (foxes, mustelids), I would often have a second person review the ID, or do it myself. Later on, I had a student go through all the tagged images of a particular species (gray squirrel, etc.) and verify the first ID. I found that some of the undergrads were very reliable in their ability to ID the species, whereas some other students needed to have their work checked more meticulously. I later thought of the idea of building a training set of say, 100 photos, to have each student run through to get a sense for their familiarity with the species, but also their ability to handle the more tricky scenarios that come up often in camtrap datasets.

Most folks could only handle 1-1.5 hours of continuous tagging. I had a few enthusiasts who would go for 2 hours straight, but that was rare. We logged effort in a shared google spreadsheet, where the students noted the dataset they worked on, any issues that came up, and any individual images that needed a second check.

I also tried to set up a more ergonomic workstation for folks (multiple monitors raised up, ergonomic mouse, etc.). Since the motion is so repetetive, easy for folks to develop carpal tunnel syndrome.

If you are dealing with a much larger dataset, you might want to look into more sophisticated AI/automation methods, but for a smaller project, this was doable. If you have a university connection, you can often recruit folks through chapter groups of The Wildlife Society. Student are often eager to gain experience, although many don't stick with it once they find out how unglamorous it is!

Good luck!

-Andy

I see. Im interested and would like to help. I will need the images to train the network. As many as possible. 
if you dont have them yet, try to find similar images preferably of the same species. I will use them to test the performance of the detection.

I'm not familiar with camera traps, but there are a couple of options:

1) If the animals tend to cover most part of the image, then you can train a CNN classifier to distinguish between species (available with the keras-Tensorflow modules in Python)

2) If, however, the animals only cover a small part of the image (e.g. in the distance), it might be better to use an object detector (I've used YOLOv2 in the past for fish detection), which however is not that straightforward, especially with Python (I used MATLAB)

In any case, keras-Tensorflow classification with Python might be the most straightforward option for your goal. You should also certainly have a look at Google's Wildlife Insights platform which is specialized for species classification from camera trap images.

Hi Andrew,

Dan here—I'm one of the authors of the TinyML book! I love your Withymbe project; I've previously done work involving embedded systems and insects, and it's interesting to hear about your plans for bird flocks.

As long as you have sufficient data, you should be able to identify different bird sounds and discern them from background noise. The TinyML book has a chapter that introduces the underlying techniques, and I'd also recommend taking a look at www.edgeimpulse.com - we've built a set of tools designed to make it easy to train these types of models.

We actually recently published a tutorial on Wildlabs about this very concept:

https://www.wildlabs.net/resources/case-studies/tutorial-train-tinyml-model-can-recognize-sounds-using-only-23-kb-ram

I'm always excited to learn about new applications; feel free to reach out if there's any way we can help. I'm [email protected].

Warmly,
Dan