Semi-supervised camouflaged fixation prediction via self-evolving pseudo-label learning

Fixation prediction in camouflaged scenes is challenging, as severe foreground–background similarity weakens discriminative cues and makes eye-tracking annotations costly and scarce. Existing fixation prediction models are typically developed for natural scenes with more distinguishable foreground–background contrast, which limits their effectiveness in camouflage settings. Moreover, annotation scarcity and the inherent ambiguity of camouflaged scenes make it difficult to obtain sufficient reliable supervision. To address these issues, we propose Camouflaged Fixation Prediction (CFP), a semi-supervised framework that serves as both a fixation prediction network and an adaptive teacher for unlabeled data. As a fixation prediction network, CFP integrates multi-scale feature encoding, explicit background estimation and suppression, and hierarchical fusion decoding to enhance fixation cues under low contrast, complex textures, and heavy occlusions. As the adaptive teacher, CFP introduces a self-evolving pseudo-label mechanism within a closed-loop teacher–student paradigm, enabling pseudo supervision to be progressively refined under camouflage-induced ambiguity. Extensive experiments on the CAM-FR dataset demonstrate the effectiveness of the proposed framework. Notably, CFP outperforms state-of-the-art fully supervised saliency models while using only 5% of the annotated training data. Code available at: https://github.com/RobertTang0/CFP.