A benchmark dataset and baseline model for co-salient object detection within RGB-D images

Within-image co-salient object detection (wCoSOD) identifies the common and salient objects within an image, which can benefit for many applications, such as reducing information redundancy, animation synthesis, and so on. Besides, the introduction of depth information that conforms to the stereo perception of human is also more conducive to accurately detecting salient objects. Thus, in this paper, we focus on a new task from the perspective of the benchmark dataset and baseline model, i.e., within-image co-salient object detection in RGB-D images. To bridge the gap the new task and algorithm verification, we first collect a new dataset containing 240 RGB-D images and the corresponding pixel-wise ground truth. Then, we propose an unsupervised method for within-image co-salient object detection in RGB-D images. Under the constraint of depth information, our model decomposes the within-image co-salient object detection task into two parts: determining the salient object proposals; combining the similarity constraint and cluster-based constraint between different proposals to locate the co-salient object and generate the final result. The experimental results on the collected dataset demonstrate that our method achieves competitive performance both qualitatively and quantitatively.