Cooperative Bi-path Metric for Few-shot Learning

Zeyuan Wang,   Yifan Zhao,   Jia Li*,  

State Key Laboratory of Virtual Reality Technology and Systems, Beihang University

Yonghong Tian

School of Electronics Engineering and Computer Science, Peking University

Given base classes with sufficient labeled samples, the target of few-shot classification is to recognize unlabeled samples of novel classes with only a few labeled samples. Novel classes and scarcity of labeled samples make few-shot classification much challenging. Most of existing methods only pay attention to the relationship between labeled and unlabeled samples of novel classes, which don’t make full use of information within base classes. In this paper, we make two contributions to investigate few-shot classification problem. First, we report a simple and effective baseline trained on base classes in the way of traditional supervised learning, which can achieve comparable results to the state of the art. Second, based on the baseline, we propose a coorperative bi-path metric for classification, which leverages the correlations between base classes and novel classes to further improve the accuracy. Experiments on two widely used benchmarks show that our method is a simple and effective framework, and a new state of the art is established in the few-shot classification field.


Illustration of the pipeline of Cooperative Bi-path Metric. The top half part upon the red dotted line represents our proposed transductive similarity, and the bottom half part is the classic inductive similarity. ρsupport refers to the similarity distributions of support set on base classes, and ρquery refers to those of query set. The final classification score ψ is the weighted sum of ϕ and φ, where ϕ is the usual cosine similarity between support set and query set, and φ is the similarity between ρsupport and ρquery.


Comparison to prior works on 5-way classification on miniImageNet benchmark. ConvNet is a 4-layer convolutional network, WideResNet is a wider version of ResNet, baseline++ denotes our baseline with tricks and CBM (CBM+LLE) indicates Coorperative Bi-path Metric (with LLE). All the numbers of prior works are imported from corresponding original papers. Best results are bolded

 Update logs

2020/08: We have updated the webpage.