STARS: Spatial-Temporal Active Re-sampling for Label-Efficient Learning from Noisy Annotations

Active learning (AL) aims to sample the most informative data instances for labeling, which makes model fitting efficient while significantly reducing annotation cost. However, existing AL models make a strong assumption that annotated are always assigned correct labels, may not hold true in many practical settings. In this paper, we develop theoretical framework formally analyze impact of noisy annotations and show systematically re-sampling guarantees reduce noise rate, can lead improved generalization capability. More importantly, demonstrates key benefit conducting active on label-efficient learning, is critical AL. The results also suggest essential properties an function with fast convergence speed guaranteed error reduction. This inspires us design novel spatial-temporal by leveraging important spatial temporal maximum-margin classifiers. Extensive experiments conducted both synthetic real-world clearly demonstrate effectiveness proposed function.