Abstract
As an important and challenging problem, image generation with limited data aims at generating realistic images through training a GAN model given few samples. A typical solution is to transfer a well-trained GAN model from a data-rich source domain to the data-deficient target domain. In this paper, we propose a novel self-supervised transfer scheme termed D3T-GAN, addressing the cross-domain GANs transfer in limited image generation. Specifically, we design two individual strategies to transfer knowledge between generators and discriminators, respectively. To transfer knowledge between generators, we conduct a data-dependent transformation, which projects target samples into the latent space of source generator and reconstructs them back. Then, we perform knowledge transfer from transformed samples to generated samples. To transfer knowledge between discriminators, we design a multi-level discriminant knowledge distillation from the source discriminator to the target discriminator on both the real and fake samples. Extensive experiments show that our method improves the quality of generated images and achieves the state-of-the-art FID scores on commonly used datasets.
- [1] . 2019. Image2StyleGAN: How to embed images into the StyleGAN latent space?. In Proc. ICCV. 4432–4441.Google Scholar
- [2] . 2017. Data augmentation generative adversarial networks. In Proc. ICLR.Google Scholar
- [3] . 2017. Wasserstein generative adversarial networks. In Proc. ICML. PMLR, 214–223.Google Scholar
- [4] . 2018. Few-shot generative modelling with generative matching networks. In Proc. AISTATS. PMLR, 670–678.Google Scholar
- [5] . 2019. Large scale GAN training for high fidelity natural image synthesis. In Proc. ICLR.Google Scholar
- [6] . 2019. FIGR: Few-shot image generation with reptile. In arXiv preprint arXiv:1901.02199.Google Scholar
- [7] . 2016. Generating images with perceptual similarity metrics based on deep networks. In Proc. NeurIPS, Vol. 29. 658–666.Google Scholar
- [8] . 2021. Federated few-shot learning with adversarial learning. In arXiv preprint arXiv:2104.00365.Google Scholar
- [9] . 2020. A pairwise attentive adversarial spatiotemporal network for cross-domain few-shot action recognition-R2. In IEEE Trans. Image Process., Vol. 30. IEEE, 767–782.Google Scholar
- [10] . 2014. Generative adversarial networks. In Proc. NeurIPS.Google Scholar
- [11] . 2020. Image processing using multi-code GAN prior. In Proc. CVPR. 3012–3021.Google Scholar
- [12] . 2017. Improved training of Wasserstein GANs. In Proc. NeurIPS.Google Scholar
- [13] . 2020. Few-shot fish image generation and classification. In Global Oceans. IEEE, 1–6.Google Scholar
- [14] . 2020. Sketch-guided deep portrait generation. In ACM Trans. Multimedia Comput. Commun. Appl., Vol. 16. ACM New York, NY, USA, 1–18.Google Scholar
- [15] . 2020. MatchingGAN: Matching-based few-shot image generation. In Proc. IEEE ICME. IEEE, 1–6.Google Scholar
- [16] . 2020. F2GAN: Fusing-and-filling GAN for few-shot image generation. In Proc. ACM MM. 2535–2543.Google Scholar
- [17] . 2017. Arbitrary style transfer in real-time with adaptive instance normalization. In Proc. ICCV. 1501–1510.Google Scholar
- [18] . 2020. Few-shot food recognition via multi-view representation learning. In ACM Trans. Multimedia Comput. Commun. Appl., Vol. 16. ACM New York, NY, USA, 1–20.Google Scholar
- [19] . 2018. Progressive growing of GANs for improved quality, stability, and variation. In Proc. ICLR.Google Scholar
- [20] . 2020. Training generative adversarial networks with limited data. In Proc. NeurIPS.Google Scholar
- [21] . 2019. A style-based generator architecture for generative adversarial networks. In Proc. CVPR. 4401–4410.Google Scholar
- [22] . 2020. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR. 8110–8119.Google Scholar
- [23] . 2015. Adam: A method for stochastic optimization. In Proc. ICLR.Google Scholar
- [24] . 2019. Asymmetric GAN for unpaired image-to-image translation. In IEEE Trans. Image Process., Vol. 28. IEEE, 5881–5896.Google Scholar
- [25] . 2020. Few-shot image generation with elastic weight consolidation. In Proc. NeurIPS.Google Scholar
- [26] . 2020. DAWSON: A domain adaptive few shot generation framework. In arXiv preprint arXiv:2001.00576.Google Scholar
- [27] . 2020. Data instance prior for transfer learning in GANs. In arXiv preprint arXiv:2012.04256.Google Scholar
- [28] . 2018. Which training methods for GANs do actually converge?. In Proc. ICML. PMLR, 3481–3490.Google Scholar
- [29] . 2018. Spectral normalization for generative adversarial networks. In Proc. ICLR.Google Scholar
- [30] . 2020. Freeze discriminator: A simple baseline for fine-tuning GANs. In Proc. CVPR Workshop.Google Scholar
- [31] . 2018. On first-order meta-learning algorithms. In arXiv preprint arXiv:1803.02999.Google Scholar
- [32] . 2019. Image generation from small datasets via batch statistics adaptation. In Proc. ICCV. 2750–2758.Google Scholar
- [33] . 2016. F-GAN: Training generative neural samplers using variational divergence minimization. In Proc. NeurIPS.Google Scholar
- [34] . 2021. Fast adaptive meta-learning for few-shot image generation. In IEEE Trans. Multimedia. IEEE.Google Scholar
- [35] . 2016. Unsupervised representation learning with deep convolutional generative adversarial networks. In Proc. ICLR.Google Scholar
- [36] . 2020. Few-shot adaptation of generative adversarial networks. In arXiv preprint arXiv:2010.11943.Google Scholar
- [37] . 2011. Learning hybrid image templates (HIT) by information projection. In IEEE Trans. Pattern Anal. Mach. Intell., Vol. 34. IEEE, 1354–1367.Google Scholar
- [38] . 2016. Rethinking the inception architecture for computer vision. In Proc. CVPR. 2818–2826.Google Scholar
- [39] . 2020. Unified generative adversarial networks for controllable image-to-image translation. In IEEE Trans. Image Process., Vol. 29. IEEE, 8916–8929.Google Scholar
- [40] . 2021. On data augmentation for GAN training. In IEEE Trans. Image Process.IEEE, 1882–1897.Google Scholar
- [41] . 2019. Few-shot video-to-video synthesis. In Proc. NeurIPS.Google Scholar
- [42] . 2019. U-Net conditional GANs for photo-realistic and identity-preserving facial expression synthesis. In ACM Trans. Multimedia Comput. Commun. Appl., Vol. 15. ACM New York, NY, USA, 1–23.Google Scholar
- [43] . 2020. MineGAN: Effective knowledge transfer from GANs to target domains with few images. In Proc. CVPR. 9332–9341.Google Scholar
- [44] . 2018. Transferring GANs: Generating images from limited data. In Proc. ECCV. 218–234.Google Scholar
- [45] . 2019. Few-shot learning: A survey. In arXiv preprint arXiv:1904.05046.Google Scholar
- [46] . 2021. Data-efficient instance generation from instance discrimination. In arXiv preprint arXiv:2106.04566.Google Scholar
- [47] . 2015. LSUN: Construction of a large-scale image dataset using deep learning with humans in the loop. In arXiv preprint arXiv:1506.03365.Google Scholar
- [48] . 2019. Self-attention generative adversarial networks. In Proc. ICML. PMLR, 7354–7363.Google Scholar
- [49] . 2019. Consistency regularization for generative adversarial networks. In Proc. ICLR.Google Scholar
- [50] . 2017. Energy-based generative adversarial network. In Proc. ICLR.Google Scholar
- [51] . 2020. On leveraging pretrained GANs for generation with limited data. In Proc. ICML. PMLR, 11340–11351.Google Scholar
- [52] . 2020. Differentiable augmentation for data-efficient GAN training. In Proc. NeurIPS.Google Scholar
- [53] . 2021. Improved consistency regularization for GANs. In Proc. AAAI.Google Scholar
- [54] . 2020. Image augmentations for GAN training. In arXiv preprint arXiv:2006.02595.Google Scholar
- [55] . 2020. In-domain GAN inversion for real image editing. In Proc. ECCV. Springer, 592–608.Google Scholar
- [56] . 2016. Generative visual manipulation on the natural image manifold. In Proc. ECCV. Springer, 597–613.Google Scholar
- [57] . 2021. Few-shot common-object reasoning using common-centric localization network. In IEEE Trans. Image Process., Vol. 30. IEEE, 4253–4262.Google Scholar
Index Terms
D3T-GAN: Data-Dependent Domain Transfer GANs for Image Generation with Limited Data
Recommendations
On leveraging pretrained GANs for generation with limited data
ICML'20: Proceedings of the 37th International Conference on Machine LearningRecent work has shown generative adversarial networks (GANs) can generate highly realistic images, that are often indistinguishable (by humans) from real images. Most images so generated are not contained in the training dataset, suggesting potential for ...
Transferring GANs: Generating Images from Limited Data
Computer Vision – ECCV 2018AbstractTransferring knowledge of pre-trained networks to new domains by means of fine-tuning is a widely used practice for applications based on discriminative models. To the best of our knowledge this practice has not been studied within the context of ...
Unpaired multi-domain image generation via regularized conditional GANs
IJCAI'18: Proceedings of the 27th International Joint Conference on Artificial IntelligenceIn this paper, we study the problem of multidomain image generation, the goal of which is to generate pairs of corresponding images from different domains. With the recent development in generative models, image generation has achieved great progress ...






Comments