Avoid wasted annotation costs in open-set active learning with pre-trained vision–language model

Active learning (AL) aims to improve model performance by selectively collecting highly informative data and reducing annotation costs. In practice, however, unlabeled data may include out-of-distribution (OOD) samples that are not used for training, which can lead to wasted annotation effort if such data are incorrectly selected. To make AL more practical, it is necessary to consider both informativeness and purity, that is, whether a sample belongs to the in-distribution (ID). Recent studies have attempted to incorporate purity estimation into AL frameworks. However, these approaches still rely heavily on OOD data and fail to jointly optimize informativeness and purity, resulting in suboptimal sample selection. To address these challenges, this study proposes Vision–Language-based Purity-oriented AL (VLPure-AL). The framework increases the purity of selected samples. It also minimizes annotation costs by reducing reliance on OOD data. VLPure-AL sequentially evaluates the purity and informativeness of data. First, it utilizes a pre-trained vision–language model to detect and exclude OOD data with high accuracy by leveraging linguistic and visual information of ID data. Second, it selects highly informative data from the remaining ID data, and then the selected samples are annotated by human experts. Experimental results on datasets with various open-set conditions demonstrate that VLPure-AL achieves the lowest cost loss and highest performance across all scenarios.