Potential of Diffusion-Generated Data on Salient Object Detection

With the development of deep learning, salient object detection (SOD) has made significant progress. However, this advancement is often constrained by the requirement for extensive training data and expensive manual annotation. To eliminate the laborious cost of dataset collection and pixel-level annotation, in this work, we employ Stable Diffusion to synthesize data and subsequently automate annotation for the SOD task. Firstly, we design a unified prompt and ChatGPT4 driven diverse prompts, which guide generating images with simple and complex scenes using Stable Diffusion. Secondly, the reliable pseudo-labels of these synthetic images are generated. For simple images, we propose the simple pseudo-label generation (SPLG) strategy which combines SAM segmentation and CLIP classifier, then train the initial SOD model. For complex images, we utilize the inference capability of the initial SOD model to generate pseudo-labels using the complex pseudo-label generation (CPLG) strategy, and employ iterative training to dynamically update the pseudo-labels. Finally, we design a simple yet effective SOD model which combines a feature fusion module (FFM) and an edge enhancement module (EEM), the former is employed to extract saliency via fusing high-level features, and the latter extracts spatial positional information from low-level features to enhance the edges of saliency results. Experiments on five benchmarks show that our method outperforms the unannotated methods, and also demonstrates better or comparable performance than weak annotation based methods. Our code will be published at https://github.com/FangWenRE/PotentialOfSDSOD.