Fine-Grained Partial Label Learning

Partial Label Learning (PLL) involves associating each instance with a candidate label set, which includes the true label along with false positive labels. Traditional PLL approaches typically assume that these false positive labels follow either a uniform or a non-uniform structure, based on the true label of an instance. However, these assumptions do not fully encompass all aspects of real-world scenarios, especially in cases of finegrained objects with subtle differences in appearance. Therefore, we introduce a fine-grained PLL task specifically designed to handle candidate sets that comprise the true label, partially nonuniform fine-grained labels manifesting in a hierarchical structure, and other false positive labels. Nevertheless, the incorporation of these partially non-uniform, fine-grained labels can result in an intractable posterior, leading to an inconsistent classifier. To address this issue, we propose a Specific Classes-Hierarchical (SCH) regularisation that ensures classifier consistency. Subsequently, the Global Label-Hierarchical-wise Embedding (GLHE) regularisation is introduced, using the refined pseudo labels of positive samples from the obtained consistent classifier. This allows the model to learn more distinctive representations from finegrained instances. Our extensive experiments on datasets such as CIFAR-10, CIFAR-100, and CUB-200 have demonstrated the effectiveness of our approach.