Causality-Driven Debiased Prototype Learning for Few-Shot Semantic Segmentation

Prototype-based approaches have emerged as a dominant method in few-shot semantic segmentation, where prototype representations serve as class-level embeddings to transfer knowledge from support to query samples. However, their effectiveness on novel classes is often undermined by the unreliability of prototypes constructed from limited support data. To alleviate this challenge, we propose CauDPL, a causality-driven prototype learning framework that examines the confounding effects of pre-trained knowledge on prototype extraction, thereby enhancing prototype robustness. Specifically, CauDPL constructs a structural causal model (SCM) to explicitly represent the causal relationships among pre-trained knowledge, sample features, and segmentation predictions, allowing for principled reasoning about how shared features across different classes influence prototype reliability. Based on this SCM, CauDPL applies causal intervention via backdoor adjustment to mitigate the confounding effects of global base-category semantic information, thereby reducing spurious correlations between sample features and class labels in few-shot scenarios and enhancing the generalization of prototypes to novel classes. Experiments on two widely-used datasets demonstrate the effectiveness of CauDPL in reducing biases and its applicability to other state-of-the-art methods.