One Coin Has Two Sides: Single Poistive Multi Label Learning from Salient Annotations

Single-Positive Multi-Label Learning (SPML) studies learning from incomplete supervision, where each instance is annotated with only one positive label despite potentially belonging to multiple categories. While existing methods assume the annotated labels are randomly distributed, real-world annotations are often biased toward the most salient category. We formalize this realistic scenario as Salient Single-Positive Multi-Label Learning (SalSPML). This salient annotation bias poses a challenge to conventional SPML methods, as the missing labels often correspond to less salient and harder-to-recognize categories. Fortunately, we find that salient annotations are typically more representative and informative. Motivated by this insight, we propose Prototype-Guided Rejection for Salient Annotation (PiSA), which constructs reliable class-wise prototypes from salient labels and leverages them to guide embedding learning for non-salient labels recognition. We theoretically demonstrate that SalSPML is harder than Random SPML due to irreducible annotation bias, and under SalSPML, more accurate prototypes facilitate false-negative label detection. Experiments on multiple benchmarks, together with two newly constructed real-world SalSPML datasets, demonstrate that PiSA consistently outperforms existing methods, achieving an average mAP improvement of 3.16\%.

Many real-world images contain multiple objects, but labeling every object is expensive and time-consuming. In practice, people often annotate only the most obvious object, such as a large or central one, while smaller or less visible objects are left unlabeled. This makes learning difficult, because a model may wrongly treat these missing objects as absent. In this work, we study this realistic labeling situation and show that it is harder than the commonly assumed setting where one label is chosen at random. We also observe that the objects people choose to label are often clear and representative examples of their categories. Based on this insight, we propose PiSA, a method that uses these obvious labeled objects as reliable references for each category. The model then uses these references to better identify objects that were present but not labeled. Our experiments show that PiSA improves recognition under incomplete and saliency-biased human annotations, making multi-label image learning more reliable when only limited labels are available.