Introduction to "classical" Animal Classifiers: How They Work and What to Expect

When it comes to identifying animals in images, there's a fascinating world of technology behind the scenes. In this article, I'll delve into the theory behind how many animal classifiers work, share some helpful resources, and set the stage for exploring results and experiences in my next post.

Before diving in, a quick note: In the project I’m working on with my team Animal Detect we’re exploring alternative approaches to classification that I won’t cover here. This article focuses on the more traditional pipelines commonly used in animal detection.

The Typical Animal Classification Pipeline

The animal classification process often follows a well-defined set of steps. Here's an overview:

Step 1: Use a “Binary” Detector

The first step is finding a general-purpose detector that identifies whether animals are present in an image, without worrying about specific species or classes.

A great example for land-based trail cameras is MegaDetector, which excels at identifying animals broadly. For underwater use, tools like MegaFishDetector serve a similar purpose.

Step 2: Crop Out the Animal

Once the detector has identified an animal, the next step is to crop out the relevant portion of the image containing the animal. This allows the classifier to focus only on the subject.

Step 3: Classify the Animal

The cropped image is then passed through a convolutional neural network (CNN) for classification. Popular models like ResNet18 or ResNet50 are commonly used in these pipelines due to their balance of accuracy and efficiency.

Here’s a helpful visual from DeepFaune that summarizes these steps perfectly:

Train Your Own Classifier

If you’re interested in customizing a classifier for your own needs, the team maintinaing MegaDetector provides a framework through Pytorch-Wildlife. You can train your own model using your dataset by following the steps outlined in their GitHub repository.

Use Pre-Trained Classifiers

There are also several pre-trained classifiers available online, often hidden deep in GitHub repositories. These models typically focus on specific regions or groups of animals, such as one model for nine African animals or another for 16 European animals (Numbers only as an example). Unfortunately, there’s no one-size-fits-all solution... yet, but they can be a great starting point.

One standout option is EcoAssist by Addax Data Science, which I’ve mentioned in a previous post. EcoAssist provides an open-source tool with nine different models for animal classification. Check them out here: EcoAssist

Quick notes about EcoAssist:

• It’s relatively easy to use, though I’ve run into errors while setting it up. Once configured, it works fine.
• All processing happens locally on your computer, so a powerful PC is necessary for speedy results.

In my next article, I’ll dive deeper into testing classifiers from EcoAssist and sharing results—so stay tuned for that! ??

Final Thoughts

Animal classification technology has come a long way, offering exciting possibilities for researchers and enthusiasts alike. Whether you’re training your own classifier or using pre-trained models, there’s a lot to explore and experiment with. But with new tools and technologies emerging quickly, there may be something pushing the wildlife conservation technology further!

I hope you found this introduction insightful! Feel free to share your thoughts or ask questions—I’d love to hear about your experiences with animal classifiers.