Multi-level Second-order Few-shot Learning

We propose a Multi-level Second-order (MlSo) few-shot learning network for supervised or unsupervised image classification and action recognition. leverage so-called power-normalized second-order base learner streams combined with features that express multiple levels of visual abstraction, we use self-supervised discriminating mechanisms. As Pooling (SoP) is popular in recognition, employ its basic element-wise variant our pipeline. The goal multi-level feature design to extract representations at different layer-wise CNN, realizing several abstraction achieve robust learning. SoP can handle convolutional maps varying spatial sizes, also introduce inputs scales into MlSo. To exploit the discriminative information from multi-scale features, develop Feature Matching (FM) module reweights their respective branches. step, which discriminator level scale abstraction. Our pipeline trained an end-to-end manner. With simple architecture, demonstrate respectable results on standard datasets such as Omniglot, mini-ImageNet, tiered-ImageNet, Open MIC, fine-grained CUB Birds, Stanford Dogs Cars, recognition HMDB51, UCF101, mini-MIT.