SML: Semantic meta-learning for few-shot semantic segmentation?

The significant amount of training data required for Convolutional Neural Networks has become a bottleneck applications like semantic segmentation. Few-shot segmentation algorithms address this problem, with an aim to achieve good performance in the low-data regime, few annotated images. Recent approaches based on class-prototypes computed from available have achieved immense success task. In work, we propose novel meta-learning framework, Semantic Meta-Learning (SML), which incorporates class level descriptions generated prototypes problem. addition, use well-established technique, ridge regression, not only bring class-level information, but also effectively utilise information multiple images present prototype computation. This simple closed-form solution, and thus can be implemented easily efficiently. Extensive experiments benchmark PASCAL-5i dataset under different experimental settings demonstrate effectiveness proposed framework.