Co-saliency detection (Co-SOD) aims to segment the common salient foregroundin a group of relevant images. In this paper, inspired by human behavior, wepropose a gradient-induced co-saliency detection (GICD) method. We firstabstract a consensus representation for the grouped images in the embeddingspace; then, by comparing the single image with consensus representation, weutilize the feedback gradient information to induce more attention to thediscriminative co-salient features. In addition, due to the lack of Co-SODtraining data, we design a jigsaw training strategy, with which Co-SOD networkscan be trained on general saliency datasets without extra pixel-levelannotations. To evaluate the performance of Co-SOD methods on discovering theco-salient object among multiple foregrounds, we construct a challenging CoCAdataset, where each image contains at least one extraneous foreground alongwith the co-salient object. Experiments demonstrate that our GICD achievesstate-of-the-art performance. Our codes and dataset are available athttps://mmcheng.net/gicd/.