Low-shot Visual Recognition by Shrinking and Hallucinating Features

ICCV 2017  ·  Bharath Hariharan, Ross Girshick ·

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.

PDF Abstract ICCV 2017 PDF ICCV 2017 Abstract

Datasets

ImageNet
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Few-Shot Image Classification ImageNet-FS (1-shot, novel) SGM (ResNet-50) Top-5 Accuracy (%) 52.9 # 7
Compare
Few-Shot Image Classification ImageNet-FS (2-shot, novel) SGM w/G (ResNet-50) Top-5 Accuracy (%) 64.9 # 8
Compare
Few-Shot Image Classification ImageNet-FS (2-shot, novel) SGM (ResNet-50) Top-5 Accuracy (%) 67.0 # 6
Compare
Few-Shot Image Classification ImageNet-FS (5-shot, all) SGM (ResNet-50) Top-5 Accuracy (%) 77.4 # 5
Compare
Few-Shot Image Classification ImageNet-FS (5-shot, all) SGM w/ G (ResNet-50) Top-5 Accuracy (%) 77.3 # 7
Compare

Methods
Add Remove

No methods listed for this paper. Add relevant methods here
Contact us on: hello@paperswithcode.com . Papers With Code is a free resource with all data licensed under CC-BY-SA.
Terms Data policy Cookies policy