Improving Landmark Localization with Semi-Supervised Learning
share
We present two techniques to improve landmark localization from partially annotated datasets. Our primary goal is to leverage the common situation where precise landmark locations are only provided for a small data subset, but where class labels for classification tasks related to the landmarks are more abundantly available. We propose a new architecture for landmark localization, where training with class labels acts as an auxiliary signal to guide the landmark localization on unlabeled data. A key aspect of our approach is that errors can be backpropagated through a complete landmark localization model. We also propose and explore an unsupervised learning technique for landmark localization based on having a model predict equivariant landmarks with respect to transformations applied to the image. We show that this technique, used as additional regularization, improves landmark prediction considerably and can learn effective detectors even when only a small fraction of the dataset has labels for landmarks. We present results on two toy datasets and three real datasets, with hands and faces, respectively, showing the performance gain of our method on each.
READ FULL TEXT