Animal Pose Estimation with DeepLabCut

DeepLabCut was introduced in peer-reviewed research as markerless tracking of user-defined body parts with deep learning. A researcher chooses landmarks relevant to a question, such as a nose, tail base or paw, and labels them on selected video frames. The trained model then predicts locations for those parts across more frames. Markerless means physical markers need not be attached to the animal; it does not mean the training process requires no human labels or quality checks. The result is usually a set of image-plane keypoint coordinates with a likelihood or confidence-like output. It is not automatically a full 3D body mesh, an identity label, or an interpretation of behavior. If a paw disappears behind a cage wall, a model may still output a position, but the image contains limited evidence. Review uncertain frames and consider whether an apparent jump reflects movement, occlusion or an error. The official guide shows how predicted points and their likelihoods can be inspected against human labels. The training examples should represent the animals, camera positions, lighting, backgrounds and poses expected in the study. A model that works on one recording may struggle when a cage, camera or species changes. Closely neighboring frames from the same clip are not independent evidence of generalization; evaluate on separate sequences or conditions where possible. Check part-specific error and failure cases, not just an overall average. In multi-animal scenes, locating points and assigning them to the right individual are related but separate problems. Researchers can use trajectories to quantify defined movements, but a nose path does not by itself reveal an animal's intention, pain or emotional state. Define the behavioral measure and validate it against independent observations. For 3D pose, multiple synchronized and calibrated camera views add depth evidence; a single 2D view cannot uniquely recover hidden depth. Preserve study metadata and version the labels, model and processing settings so results can be reproduced and corrections traced.