Instance-Proxy Loss for Semi-supervised Learning with Coarse Labels

Objects are often organized in a hierarchy where coarse-grained categories are comprised of subordinate fine-grained classes. Comparing with the fine-grained labels, the coarse-grained labels are much affordable to obtain. The coarse-grained labels can boost the semi-supervised learning (SSL) by offering extra regularization on the feature space of finer-grained recognition. However, coarse-grained labels are ignored by most of works in SSL. An intuitive way to utilize the coarse labels for SSL is to impose an extra coarse-grained categorization constraint, which will cause the class confusion between fine-grained categories belonging to the same coarse-grained category thus is sub-optimal for SSL. In this paper, we present an instance-proxy loss (IPL) to boost the separability of the fine-grained classes within the same coarse-grained class, as well as keep the intra-class feature space of coarse-grained classes compact. Specifically, IPL includes instance-level loss and proxy-level loss to impose constraints on both instance-to-instance and instance-to-proxy relations. Our approach outperforms the state-of-the-art methods on three benchmark datasets, showing significant improvement with small proportion of fine-grained labels, e.g., it brings 10.14% accuracy improvement on CUB-200-2011 with 15% of labeled data.