Learning to Predict Anthropometric Landmarks via Feature Refinement

Precise localization of anthropometric landmarks is essential in many applications ---including computer graphics, computer vision, biomechanics, and morphometric studies.Current methods in machine learning can produce dense correspondence with good global properties, but are poor at localizing specific landmarks.Landmarks are the most important correspondences for many applications and the only independently verifiable criteria.Here we propose a new method for localizing landmarks using learned features that significantly outperforms the state of the art.Our method learns refined features that characterize the intrinsic and extrinsic geometry around the landmark, thereby making it easy to recognize and localize.We propose a novel loss function for training using two functions to evaluate the likelihood of a vertex being a target landmark: the landmark potential ( ), which characterizes the likelihood of a vertex being a landmark and is computed by the network from a given mesh, and the similarity function ( ), which measures the distance from a landmark, extrinsically or/and intrinsically, and provides supervision information for training.The network is trained to optimize the correlation between the two functions so that vertices that have high similarity with the target landmark also have high potential, and vice versa.In addition, we introduce nuclear-norm minimization to compute an optimal span of refined features for landmarks among training models.The resulting system is easy to implement and learns efficiently from a small number of meshes with identified landmarks.Our implantation is available at https://github.com/yibojiao211/Learning_to_Predict_Landmarks.