Salient object detection (SOD) focuses on identifying the most prominent region in an image. Deep learning based SOD methods usually need a substantial number of images with full or weak annotation. To alleviate the high cost of annotation, in this paper, we present a SOD method without any annotation. Firstly, we propose one Spatial-enhanced Contrastive learning for Class-agnostic Activation Map (SCCAM). Given each training image without any annotation, SCCAM can predict activation map which highlights the approximate locations of the objects. Secondly, point prompts and box prompts are generated from the activation maps and inputted into Segment Anything (SAM) to produce two segmentation results, then generate initial reliable pseudo-labels and training initial SOD model. Thirdly, we use the inference from initial SOD model as guidance to generate refined reliable pseudo-labels and train refined SOD model. Experiments on four public datasets show our work outperforms other annotation-free models, and achieves comparable or superior performance than weak annotation based methods.