To extract behavioral information from rich video recordings, two avenues of research relying on deep learning have been proposed: landmark-based pose estimation methods [22,26,32] and autoencoder-based dimensionality reduction methods [2]. Pose estimation methods characterize animal behavior with the trajectories of body-part landmarks.

However,two ofthe most popular benchmarks donot provide the full training information for eacharchitecture.

However,two ofthe most popular benchmarks donot provide the full training information for eacharchitecture.