Self-Supervised Keypoint Discovery in Behavioral Videos

We propose a method for learning the posture and structure of agents from unlabelled behavioral videos. Starting from the observation that behaving agents are generally the main sources of movement in behavioral videos, our method uses an encoder-decoder architecture with a geometric bottleneck to reconstruct the difference between video frames. By focusing only on regions of movement, our approach works directly on input videos without requiring manual annotations, such as keypoints or bounding boxes. Experiments on a variety of agent types (mouse, fly, human, jellyfish, and trees) demonstrate the generality of our approach and reveal that our discovered keypoints represent semantically meaningful body parts, which achieve stateof-the-art performance on keypoint regression among selfsupervised methods. Additionally, our discovered keypoints achieve comparable performance to supervised keypoints *Equal contribution. on downstream tasks, such as behavior classification, suggesting that our method can dramatically reduce the cost of model training vis-a-vis supervised methods.