Unsupervised object segmentation based on multi-background feature similarity

The self-supervised Transformers have provided new technical approaches for unsupervised object segmentation tasks by leveraging scene modeling capabilities. Existing methods primarily rely on the similarity of self-supervised features to obtain object masks. However, due to the absence of object-level semantic information in self-supervised features, there exist feature points with high similarity between objects and backgrounds. Consequently, global similarity-based methods face challenges in effectively distinguishing these feature points, resulting in reduced segmentation accuracy. To address this challenge, we propose an unsupervised segmentation method based on multi-background similarity comparisons. This method capitalizes on the low attention of self-supervised models toward backgrounds to construct an initial background set, then performs object-background segmentation by leveraging similarity differences between object features and the initial background points. Evaluations on the DUTS dataset and aircraft segmentation dataset demonstrate that our method achieves MIoU improvements of 2.2% and 0.64% over state-of-the-art approaches, respectively. The proposed method significantly reduces misclassification errors caused by high inter-object similarity while maintaining computational efficiency.