Low-shot Visual Recognition by Shrinking and Hallucinating Features

TL;DR

This paper introduces a new low-shot learning benchmark and proposes regularization and hallucination techniques to enhance convolutional networks' ability to recognize novel classes from few examples, significantly improving accuracy.

Contribution

It presents novel regularization and hallucination methods for low-shot learning, along with a benchmark that mimics real-world recognition challenges.

Findings

Improved one-shot accuracy by 2.3x on ImageNet

Effective data augmentation for low-shot classes

Benchmark reflecting real-world recognition challenges

Abstract

Low-shot visual learning---the ability to recognize novel object categories from very few examples---is a hallmark of human visual intelligence. Existing machine learning approaches fail to generalize in the same way. To make progress on this foundational problem, we present a low-shot learning benchmark on complex images that mimics challenges faced by recognition systems in the wild. We then propose a) representation regularization techniques, and b) techniques to hallucinate additional training examples for data-starved classes. Together, our methods improve the effectiveness of convolutional networks in low-shot learning, improving the one-shot accuracy on novel classes by 2.3x on the challenging ImageNet dataset.