Self-Regularized Prototypical Network for Few-Shot Semantic Segmentation

The deep CNNs in image semantic segmentation typically require a large number of densely-annotated images for training and have difficulties generalizing to unseen object categories. Therefore, few-shot has been developed perform with just few annotated examples. In this work, we tackle the using self-regularized prototypical network (SRPNet) based on prototype extraction better utilization support information. proposed SRPNet extracts class-specific representations from generates masks query by distance metric - fidelity. A direct yet effective regularization set is SRPNet, which generated prototypes are evaluated regularized itself. extent restore mask imposes an upper limit performance. performance should never exceed no matter how complete knowledge generalized set. With specific regularization, fully exploits offers high-quality that representative each class meanwhile discriminative different classes. further improved iterative inference (IQI) module combines prototypes. Our achieves new state-of-art 1-shot 5-shot benchmarks.