Learning Robust Low-Rank Approximation for Crowdsourcing on Riemannian Manifold

Recently, crowdsourcing has attracted substantial research interest due to its efficiency in collecting labels for machine learning and computer vision tasks. This paper proposes a Rieman-nian manifold optimization algorithm, ROLA (Robust Low-rank Approximation), to aggregate the labels from a novel perspective. Specifically, a novel low-rank approximation model is proposed to capture underlying correlation among annotators meanwhile identify annotator-specific noise. More significantly, ROLA defines the label noise in crowdsourcing as annotator-specific noise, which can be well regularized by l2,1-norm. The proposed ROLA can improve the aggregation performance when compared with state-of-the-art crowdsourcing methods.