Learning to Detect Instance-Level Salient Objects Using Complementary Image Labels

Existing salient instance detection (SID) methods typically learn from pixel-level annotated datasets. In this paper, we present the first weakly-supervised approach to SID problem. Although weak supervision has been considered in general saliency detection, it is mainly based on using class labels for object localization. However, non-trivial use only instance-aware information, as instances with high semantic affinities may not be easily separated by labels. As subitizing information provides an instant judgement number of items, naturally related detecting and help separate same while grouping different parts instance. Inspired observation, propose We a novel network three branches: Saliency Detection Branch leveraging consistency locate candidate objects; Boundary exploiting discrepancy delineate boundaries; Centroid detect centroids. This complementary then fused produce map. To facilitate learning process, further progressive training scheme reduce label noise corresponding learned model, via reciprocating model prediction refreshing. Our extensive evaluations show that proposed method plays favorably against carefully designed baseline adapted tasks.