Low-shot Visual Recognition by Shrinking and Hallucinating Features
About
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.
Bharath Hariharan, Ross Girshick• 2016
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Image Classification | ImageNet 1K Challenge (novel classes) | Top-5 Acc78.5 | 110 | |
| Few-shot classification | CUB | Accuracy81.1 | 96 | |
| Generalized Few-Shot Learning | ImageNet 2012 (Novel classes) | Top-5 Accuracy76.5 | 70 | |
| Few-shot Image Classification | ImageNet FS (novel) | Top-5 Acc0.82 | 59 | |
| Low-shot Image Classification | ImageNet 1k (novel classes) | Top-5 Acc79.5 | 57 | |
| Generalized Few-Shot Learning | ImageNet All classes 2012 | Top-5 Accuracy78.5 | 50 | |
| Generalized Few-Shot Learning | AWA2 | Accuracy87.8 | 48 | |
| Few-shot Image Classification | ImageNet FS (all) | Top-5 Acc83.6 | 44 | |
| Low-shot Image Classification | ImageNet-1k (val) | Top-5 Accuracy85.2 | 40 | |
| Image Classification | Imagenet FS (All classes) | Top-5 Acc77.5 | 30 |
Showing 10 of 22 rows