DeepLabCut is a software package for animal pose estimation from videos. What makes it so powerful is the ability to use the package to retrain a specific model for your own needs. On top of this, DeepLabCut offers a well documented protocol as well as an image labeler tool that makes the process much easier. By using transfer learning, DeepLabCut allows you to achieve high landmark tracking performance by using a relatively small number of images.

The range of possible applications is huge. One of the main challenges at the time of this research was tracking facial landmarks in thermal videos and extracting the temperarures from specific areas without placing markers on the faces of the participants. One of the main advantages of using DeepLabCut is that it provides a clear, easy to use workflow that guides the user through the process of using transfer learning in order to achieve the best results on their datasets. This process also ensures that good results can be achieved with relatively little manual labeling of the training datasets. The result presented in the video below were achieved after labeling only 20 frames per participant with 7 facial landmarks. The model was initialized with a pre-trained neural network (ResNet-50) and the model was trained for 210000 iterations.

Facial landmark tracking example using the DeepLabCut package