(The image above links to the paper.)
Behavioral analysis is a crucial experimental approach in biology, and accurate quantification of behavior is especially indispensable for understanding the brain.
Traditionally, high-accuracy pose estimation has been achieved by attaching markers (such as sensors) to the subject (a person or a laboratory animal).
However, sensors are a nuisance to the subject, and they can alter behavior or place constraints on it.
One marker-free alternative is to apply a skeleton model, but developing such a model is time-consuming and has to be done on a large scale.
There are also systems that estimate pose directly from images, but these generate enormous amounts of data, which again makes them very difficult for a single laboratory to run given equipment requirements.
With DeepLabCut, anyone with just a single computer can perform pose estimation, and do so very easily.
Using image recognition based on deep learning, it works marker-free and can recognize and track whatever body parts you choose.
Because the labeling is done by deep learning, you can analyze recorded videos without having to correct them beforehand.
In other words, tracking works even if the lighting across the field you want to analyze is uneven, or if the image is somewhat distorted depending on the viewing angle.
The page of the laboratory that developed DeepLabCut is here.
The GitHub repository is here.
From actually using DeepLabCut, I found that the computer used for analysis needs a GPU (graphics card) with at least 8 GB of memory.
Among gaming PC GPUs, that means a mid- to high-end card, so comfortable analysis is not possible on an everyday laptop.
That said, even without such a high-spec machine, you can use Google Colab to access a virtual GPU and run the analysis.
I haven’t tried it myself yet, though.
In actual analyses, I found the plotting accuracy to be very high.
However, it often fails when the background differs from that of the training data, so I felt it is best to prepare training data that closely matches the environment you actually want to analyze.
The paper showed that plotting accuracy improves when you label not only the few points you are interested in, but also the overall body—for example, a rough skeleton of the subject—even if those points are not used in the analysis.
Also, if the results are unsatisfactory after training and analysis, you can remove the problematic frames or add new training datasets, allowing you to actively improve the model.
DeepLabCut itself is very user-friendly and easy for anyone to use.
Why not give it a try in your own analyses?
I’d like to write about installing and using DeepLabCut in future posts.