Balanced Loss Function for Long-tailed Semi-supervised Ship Detection

Abstract

Semi-supervised learning (SSL) has significantly reduced the reliance of the ship detection network on labeled images. However, the more realistic and challenging issue of long-tailed distribution in SSL remains largely unexplored. While most existing methods address this issue at the instance level through reweighting or resampling techniques, their performance is significantly limited by their dependence on biased backbone representations. To overcome this limitation, we propose a Balanced Loss function (Bal Loss). Our approach consists of three key components. First, we introduce the BaCon Loss, which computes class-wise feature centers as positive anchors and selects negative anchors through a simple yet effective mechanism. Second, we posit an assumption that the normalized features in contrastive learning follow a mixture of von Mises-Fisher (vMF) distributions in the unit space. This assumption allows us to estimate the distribution parameters using only the first sample moment, which can be efficiently computed in an online manner across different batches. Finally, we incorporate a Jitter-Bagging module, adapted from prior literature, to provide precise localization information, thereby refining bounding box predictions. Extensive experiments demonstrate the efficacy of Bal Loss, achieving SOTA results on ship datasets with a 3.9 improvement over the baseline. Notably, our method attains an \(AP^{r}\) of 44.1 on the ShipRSImageNet dataset, underscoring its robust detection capabilities.