PROJECT TITLE :

Prior Guided Feature Enrichment Network for Few-Shot Segmentation

ABSTRACT:

Methods of semantic segmentation that have advanced to the state-of-the-art require a sufficient amount of labeled data to achieve good results and barely work on classes that have not been seen without fine-tuning. The solution that has been proposed for this issue is called few-shot segmentation, and it involves learning a model that can easily adapt to new classes with only a few labeled support samples. These frameworks still face the challenge of generalization ability reduction on unseen classes as a result of inappropriate use of high-level semantic information of training classes and spatial inconsistency between query and support targets. This challenge was posed because inappropriate use of high-level semantic information of training classes can reduce generalization ability. To address these concerns, we offer the Prior Guided Feature Enrichment Network as a potential solution (PFENet). It is comprised of two novel designs: (1) a training-free prior mask generation method that not only maintains generalization power but also improves model performance, and (2) a Feature Enrichment Module (FEM) that overcomes spatial inconsistency by adaptively enriching query features with support features and prior masks. The first design is a training-free prior mask generation method that not only retains generalization power but also improves model performance. Extensive experiments performed on PASCAL-5 I and COCO demonstrate that the proposed prior generation method and the FEM both make significant improvements to the baseline method. Our PFENet also outperforms methods that are considered to be state-of-the-art by a significant margin and does so without sacrificing efficiency. It is astonishing that even in the absence of labeled support samples, our model can generalize to the situation.