A computing system may receive an image captured by a camera, and determine based on (1) one or more regions in the image that include a portion of a body of a user, and (2) a camera pose of the camera, a three-dimensional volume constraint that stems from the camera pose in which an elbow of the user is likely to be located. The system may infer, based on at least the three-dimensional volume constraint, an elbow pose of the user, wherein the elbow pose includes a location associated with an elbow of the user in three-dimensional space. The system may determine the location associated with the elbow is outside the three-dimensional volume constraint, and in response to the determination, adjust the inferred elbow pose. The system may generate, based on the adjusted elbow pose, a body pose of the user that includes at least the adjusted elbow pose.