Learned Gaussian ProtoNet for improved cross-domain few-shot classification and generalization

To imitate intelligent human behaviour, computer vision introduces a fundamental task called Few-Shot learning (FSL) that carries the promise of alleviating the need for exhaustively labeled data. Using prior knowledge few-shot learning aims to learn and generalize to novel tasks containing limited examples with supervised information. Although metric-based methods demonstrated promising performance but due to the large disparity of feature distributions across domains they often fail to generalize. In this work, we propose a learned Gaussian ProtoNet model for fine-grained few-shot classification via meta-learning for both in-domain and cross-domain scenarios. Gaussian ProtoNet encoder helps to map an image into an embedding vector and Gaussian covariance matrix predicts the confidence region about individual data points. Direction and class-dependent distance metrics are adopted to estimate the distances to distinct class prototypes. Feature-wise modulated layers are embedded in the encoder to augment the feature distribution of images. The learning-to-learn approach is adopted for fine-tuning the hyper-parameters of incorporated feature-wise modulated layers for better generalization on unseen domains. Experimental results justify that our proposed model performs better than many state-of-the-art models and feature-wise modulation improves the performance under domain shifts.