摘要
基于城市神经辐射场(NeRF)场景重建的主流方法是将整个场景分割为多个区域,但在无人机航拍场景中,一旦飞行轨迹变化导致场景分布不均时,现有方法的分割效率低下,难以处理同一场景不同高度拍摄的图像细节,导致渲染质量不佳。为此,提出一种端到端的注意力引导框架。首先,利用相机位姿进行自适应空间划分,解决不均匀场景分割的效率问题;其次,从不同高度图像中提取、融合目标视图最相关的特征来增强细节,解决航拍场景高保真渲染问题。实验表明,所提模型在城市场景数据集56Leonardv、SCUTic上均取得了最佳性能,执行渲染的速度相较于目前最高效的方法提升了65%以上。
Mainstream method for scene reconstruction based on urban neural radiation field(NeRF)is to divide the entire scene into multiple regions.However,in unmanned aerial vehicle(UAV)aerial scenes,once the flight trajectory changes and the scene distribution is uneven,the segmentation efficiency of existing methods is low,making it difficult to handle the details of images taken at different heights of the same scene,resulting in poor rendering quality.Therefore,an end-to-end attention guidance framework is proposed.Firstly,utilizing camera pose for adaptive spatial partitioning to address the efficiency issue of non-uniform scene segmentation;Secondly,extracting and fusing the most relevant features of the target view from images at different heights to enhance details and solve the problem of high fidelity rendering in aerial scenes.The experiment shows that the proposed model achieved the best performance on both the urban scene datasets 56Leonardv and SCUTic,with a rendering speed improvement of over 65%compared to the most efficient methods currently available.
作者
梁书凝
谭诗瀚
戈文一
钟娟
王录涛
LIANG Shuning;TAN Shihan;GE Wenyi;ZHONG Juan;WANG Lutao(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)
出处
《软件导刊》
2026年第1期127-134,共8页
Software Guide
基金
四川省科技计划项目(2023YFG0304)
四川省重大科技专项(2022ZDZX0008)。
