Transductive Zero-Shot Learning using Cross-Modal CycleGAN
Bordes, Patrick, Zablocki, Eloi, Piwowarski, Benjamin, Gallinari, Patrick
–arXiv.org Artificial Intelligence
In Computer Vision, Zero-Shot Learning (ZSL) aims at classifying unseen classes -- classes for which no matching training image exists. Most of ZSL works learn a cross-modal mapping between images and class labels for seen classes. However, the data distribution of seen and unseen classes might differ, causing a domain shift problem. Following this observation, transductive ZSL (T-ZSL) assumes that unseen classes and their associated images are known during training, but not their correspondence. As current T-ZSL approaches do not scale efficiently when the number of seen classes is high, we tackle this problem with a new model for T-ZSL based upon CycleGAN. Our model jointly (i) projects images on their seen class labels with a supervised objective and (ii) aligns unseen class labels and visual exemplars with adversarial and cycle-consistency objectives. We show the efficiency of our Cross-Modal CycleGAN model (CM-GAN) on the ImageNet T-ZSL task where we obtain state-of-the-art results. We further validate CM-GAN on a language grounding task, and on a new task that we propose: zero-shot sentence-to-image matching on MS COCO.
arXiv.org Artificial Intelligence
Nov-13-2020
- Country:
- Asia > China
- Hong Kong (0.04)
- Europe
- France > Île-de-France
- United Kingdom > England
- Oxfordshire > Oxford (0.04)
- North America
- Canada > British Columbia
- United States
- California > Los Angeles County
- Long Beach (0.04)
- Florida (0.04)
- Hawaii > Honolulu County
- Honolulu (0.04)
- Utah > Salt Lake County
- Salt Lake City (0.04)
- California > Los Angeles County
- Asia > China
- Genre:
- Research Report (0.50)
- Technology: