Learning from hybrid labels with partial labels via hybrid-grained contrast regularization

Learning from hybrid labels is suitable for dealing with the real-world scenario, where the labels of the training dataset include fine-grained labels and coarse-grained labels. Unfortunately, obtaining fine-grained labels with strong supervision information is difficult. Compared with the more difficult-to-obtain fine-grained labels, it is labor-saving and efficient for the annotator to remove several impossible options to obtain partial labels, which also provide more information than coarse-grained labels. Nonetheless, the goal of existing methods of PLL is only to learn from partially labeled data but cannot utilize fine-grained labels, coarse-grained labels, and partial labels to train an ordinary supervised classifier. In this paper, we propose a novel setting called learning from hybrid labels with partial labels (LHLP) and propose a method to learn a classifier for identifying difficult-to-recognize images by using hybrid labels with partial labels. Specifically, we drive hybrid-grained contrast regularization (HGCR) for exploring the relationship between sample labels, which uses fine-grained label information and coarse-grained label information to disambiguate labels respectively. Extensive experiments on benchmark validate the effectiveness of the proposed method HGCR in LHLP, compare to the state-of-art methods on the CIFAR-100, Kuzushiji-MNIST and Fashion-MNIST datasets, our method achieves promising results.