Vincent Duan (段嘉铭)

Duan is a first year Master student at BUPT, major in Computer Vision, Few Shot Learning, GAN

Open Source Projects

Few-shot-Playground - PyTorch’s few-shot method implementated for the latest method. (under construction)

I am new for few-shot, can u briefly describe how do few-shot model works and why it is few-shot despite it trained by a large dataset? This question is a very interesting, however very fundamentral question. In general, differ from traditional image-classification model, few-shot model learn the ability to generalize how to classifiy between the classes, so they can correctly classify the unseen classes with a few samples. More details you can see in one of my blogs here.
SOTA data? Most method are compare on the mini-imagenet, Prototypical Network, Relation Network these classic method with 4 layer of simple Conv can achieve about 65 ~ 68% Top1 acc. Recent years, lots of new work change the backbone to Resnet 34 and Prototypical Network can achieve 74% with ResNet34. This must be emphasis, because many papers didn’t mention it, so they will lead to some mis-understand, such as GNN or some novel method improved 10 per or even 15 per acc.