节点文献
基于条件归一化流的视角合成任务及其不确定性分析
Conditional Normalizing Flow Based View Synthesis with Uncertainty Analysis
【Author】 Shen Yu;Li Yuke;Wang Fei-Yue;Kunlun Digital Technology Co.,Ltd;School of Software,Northwestern Polytechnical University;Institute of Automation,Chinese Academy of Science;
【机构】 西北工业大学软件学院; 昆仑数智科技有限责任公司; 中国科学院自动化研究所;
【摘要】 针对视角合成任务中的遮挡问题,提出基于条件归一化流的可见性计算方法,以数据驱动的方式建模沿着输入视角与目标视角间的可见性,多视角特征匹配的信噪比和相似性分别提升了0.11%和0.68%。并利用统计方法,对匹配视角合成任务中的不确定性进行了度量。实验证明了所提方法的有效性和泛化性。
【Abstract】 To address the occlusion problem in view synthesis tasks,we propose a visibility computation method based on conditional normalized flow,which models visibility along the input and target views in a data-driven manner,improving the SSIM and PSNR of multi-view feature matching as 0.68% and 0.11% respectively.Additionally,statistical methods are used to measure the uncertainty in the view synthesis task.Experiments demonstrate the effectiveness and generalizability of the proposed method.
【Key words】 Conditional Normalizing Flow; View Synthesis; Uncertainty Analysis;
- 【会议录名称】 2024中国自动化大会论文集
- 【会议名称】2024中国自动化大会
- 【会议时间】2024-11-01
- 【会议地点】中国山东青岛
- 【分类号】TP391.41
- 【主办单位】中国自动化学会