1 Title
Diffusion-based Target Sampler for Unsupervised Domain Adaptation(Zhang, Yulong, Chen, Shuhao, Zhang, Yu, Lu, Jiang)【CVPR 2023】
2 Conclusion
large domain shifts and the sample scarcity in the target domain make existing UDA methods achieve suboptimal performance. To alleviate these issues, This study propose a plug-andplay Diffusion-based Target Sampler (DTS) to generate high fidelity and diversity pseudo target samples. By introducing class-conditional information, the labels of the generated target samples can be controlled.
3 Good Sentences
1、Compared with those methods that generate an intermediate domain to interpolate
between the distributions of the source and target domains, the proposed DTS framework directly generates pseudo target samples that obey the target distribution. Instead of using adversarial training strategies, the proposed DTS framework is based on the DPM, which has better generation capabilities and is easier to converge in the training process(The advantages of Diffusion model when compared with other )
2、Although DPM has obtained superior performance in image generation, it still has the problem of slow sampling speed due to thousands of denoising steps required to generate a sample of high quality, which greatly hinders the application of DPM.(The shortcomings of DPM when try to apply)
3、Different from those GAN-based methods, the proposed diffusion-based DTS framework directly generates pseudo target samples that could obey the target distribution without adversarial training. [25] has shown that DPMs are better at covering the modes of a distribution than GANs, which well meets the needs of target data generation here. And the category and the number of samples generated can be flexibly controlled.(The innovation of this paper when compared with others)
本文提出了一种基于即插即用扩散的目标采样器(DTS)来生成高保真度和多样性的伪目标样本来解决无监督域适应(UDA)中大型域偏移和目标域中的样本稀缺的问题。具体来说,DTS是生成可以遵循目标分布的伪目标样本。这样,可以用伪目标样本来增强目标样本,从而提高UDA模型的性能。DTS将生成的目标样本和原始源样本组合为增广源域,其中使用原始源样本来抑制生成目标样本的噪声标签的影响。通过这种方式,增强源域的分布更接近目标域,这降低了域自适应(DA)的难度。请注意,所提出的DTS框架是一个即插即用模块,可以嵌入到任何现有的UDA方法中,以提高其传输性能。
如图所示,整个DTS框架分为以上三个步骤,步骤1:通过一些UDA方法获得分类器,步骤2:由步骤1中预训练的分类器分配目标样本的伪标签,并使用具有伪标签的目标样本来训练CDPM。步骤3:采用预训练的CDPM来生成目标样本,并将这些生成的目标样本与原始源样本组合作为增广源域
作为一个即插即用模块,插进去了之后还是有所提升的