CVPR Daily - 2018 - Thursday

The challenge in doing that is that they need to guarantee that the generative features are good enough to train on supervised classifiers. If it’s unstable to train, they need to propose GAN architecture that is stable and can generate features with good quality. Yongqin explains what algorithmic techniques they use to solve this: “ We proposed to use Wasserstein-GAN to stabilise the GAN training. To enforce the discriminative ability of the generated features we propose a classification loss which enforces the generated features can be correctly classified by a pre-trained classifier .” Zeynep adds that they make use of the fact that the data that is surrounding us is inherently multimodal. If you have text that accompanies images, and you don’t have enough labelled images but you have text that you can use to associate different sets of classes, then you can use this model to generate more synthetic features of those classes for which you have very few examples. She tells us: “ We show the capability of this model on ImageNet . ImageNet is one of the largest scale datasets that is available to us. It generalizes to the cases when we don’t have any label training data from some of the classes, just because our model is able to associate different classes in a conditional generative adversarial net framework. ” In terms of next steps, Zeynep says they have thought about making use of the existence of text better. At the moment, they are generating image features that correspond to text sentences or attributes, but from looking at it the other way around, they can also generate text for images. They could expand this framework to do explanations, for example, explaining scene understanding and the semantic image content. Thursday 17 Yongqin Xian … you can use this model to generate more synthetic features of those classes for which you have very few examples.

RkJQdWJsaXNoZXIy NTc3NzU=