SECNet: semantic-enhanced network with calibration for few-shot semantic segmentation

Few-shot semantic segmentation (FSS) aims to segment novel query images based on a limited number of annotated support images. Despite significant progress in FSS methods in recent years, current approaches remain prone to capturing biased targets, particularly in real-world environments. To address these challenges, we argue that incorporating semantic information can help FSS models more accurately localize targets. We propose a Semantic-Enhanced Network with Calibration (SECNet), which effectively leverages semantic information to enhance image features. These enhanced features are then utilized to learn dense feature maps. Additionally, we introduce a lightweight calibration module with few learnable parameters, ensuring that the module’s performance is minimally affected by the training data. The calibration module generates a calibration map to refine the feature maps, improving robustness. To further enhance the field, we introduce a novel benchmark featuring datasets from multiple domains, offering evaluation tasks that closely resemble real-world applications. Experimental results demonstrate that SECNet outperforms competing methods on our proposed benchmark by a decent margin.