深度结构保持的场景图像生成
计算机视觉与模式识别
2017-11-23 v2
摘要
自动生成自然场景图像的关键在于恰当地安排各种空间元素，尤其是在深度方向上。为此，我们引入了一种新颖的深度结构保持场景图像生成网络（DSP-GAN），该网络采用层次化异构架构，旨在实现深度结构保持的场景生成。所提基础设施的主干建立在霍克斯点过程之上，该过程对不同深度层之间的空间依赖关系进行建模。在每个层内，生成对抗子网络以点过程生成的层信息为条件，协同训练以生成逼真的场景组件。我们在带有标注场景图像的SUN数据集子集上对模型进行了实验，结果表明我们的模型能够生成深度逼真的自然场景图像。
引用
@article{arxiv.1706.00212,
title = {Depth Structure Preserving Scene Image Generation},
author = {Wendong Zhang and Bingbing Ni and Yichao Yan and Jingwei Xu and Xiaokang Yang},
journal= {arXiv preprint arXiv:1706.00212},
year = {2017}
}
备注
There is an error in the first paragraph in Section 4.4. Actually, we train and test another new CGAN model with the input in our model to evaluate the improvements. This error can lead readers misunderstand the improvements of our model and make the comparison unfair. Therefore, we request to withdraw the current submission and will submit a final version later