基于互联网大数据的自然场景的四季合成-计算机辅助设计与图形学学报.PDF

30 5 Vol.30 No.5 2018 5 Journal of Computer-Aided Design Computer Graphics May 2018 基于互联网大数据的自然场景的四季合成 1) 2) 1)* 程乐超 , 廖子承 , 王章野 1) (浙江大学CADCG 国家重点实验室 杭州 310058) 2) (浙江大学计算机学院 杭州 310038) (zywang@cad.zju.edu.cn) 摘 要: 针对传统的纹理合成与颜色迁移对自然场景的四季合成问题存在的失效问题, 提出了基于大规模数据驱动 的方法. 首先, 从互联网上获取关于场景地点的图像并且通过全局匹配性与局部一致性约束进行粗略筛选; 其次, 利用动态规划算法简化场景图像特征点匹配, 并通过最小化约束函数确定场景参考帧以及相应场景帧; 最后, 将相 应场景帧通过变换对齐到参考帧, 并根据梯度域方法补全对齐场景从而形成场景四季序列. 实验结果表明, 该方法 能够生成更加合理的关于场景的四季图像以及四季变化鲜明的延时摄影视频. 关键词: 四季转换; 计算延时摄影; 大数据驱动的图像合成 中图法分类号: TP391.41 DOI: 10.3724/SP.J.1089.2018.16537 Synthesizing the Four Seasons of a Scene from Large-Scale Web Images Cheng Lechao1), Liao Zicheng2), and Wang Zhangye1) * 1) (State Key Laboratory of CADCG, Zhejiang University, Hangzhou 310058) 2) (College of Computer Science, Zhejiang University, Hangzhou 310038) Abstract: Traditional color or texture transfer methods are limited for synthesizing the four-season series of a scene, or transferring a scene to a different season due to the complexity of seasonal evolution. In this paper, we propose a data-driven approach, and heavily relies on large number of images on the web. We embed care- fully mined and aligned images crawled from the web into a time series of the scene, and let the manifold of complex scene-specific seasonal evolution unfold itself from the data. Because we reconstruct the time series of the scene over an entire year, we get a loopable timelapse video for free and can extract a four-season se- ries out of it. Further on, we use the result to create novel four-season mosaicking that renders interesting visual effect. Key words: season transfer; computational timelapse; data-driven