Semantic-aware representations for unsupervised Camouflaged Object Detection

Abstract

Unsupervised image segmentation algorithms face challenges due to the lack of human annotations. They typically employ representations derived from self-supervised models to generate pseudo-labels for supervising model training. Using this strategy, the model’s performance largely depends on the quality of the generated pseudo-labels. In this study, we design an unsupervised framework to perform COD (Camouflaged Object Detection) without the need for generating pseudo-labels. Specifically, we utilize semantic-aware representations, trained in a self-supervised manner on large-scale unlabeled datasets, to guide the training process. These representations not only capturing rich contextual semantic information but also assist in refining the blurred boundaries of camouflaged objects. Furthermore, we design a framework that integrates these semantic-aware representations with task-specific features, enabling the model to perform the UCOD (Unsupervised Camouflaged Object Detection) task with enhanced contextual understanding. Moreover, we introduce an innovative multi-scale token loss function, which maintain the structural integrity of objects at various scales in the model’s predictions through mutual supervision between different features and scales. Extensive experimental validation demonstrates that our model significantly enhances the performance of UCOD, closely approaching the capabilities of state-of-the-art weakly-supervised COD models.