Group attention retention network for co-salient object detection

The co-salient object detection (Co-SOD) aims to discover common, salient objects from a group of images. With the development of convolutional neural networks, the performance of Co-SOD methods has been significantly improved. However, some models cannot construct collaborative relationships across images optimally and lack effective retention of collaborative features in the top-down decoding process. In this paper, we propose a novel group attention retention network (GARNet), which captures excellent collaborative features and retains them. First, a group attention module is designed to construct the inter-image relationships. Second, an attention retention module and a spatial attention module are designed to retain inter-image relationships for protecting them from being diluted and filter out the cluttered context during feature fusion, respectively. Finally, considering the intra-group consistency and inter-group separability of images, an embedding loss is additionally designed to discriminate between real collaborative objects and distracting objects. The experiments on four datasets (iCoSeg, CoSal2015, CoSoD3k, and CoCA) show that our GARNet outperforms previous state-of-the-art methods. The source code is available at https://github.com/TJUMMG/GARNet .