侯立培,陈壹华.基于神经辐射场的室外去遮挡快速重建算法[J]. 微电子学与计算机,2024,41(5):31-39. doi: 10.19304/J.ISSN1000-7180.2023.0290
引用本文: 侯立培,陈壹华.基于神经辐射场的室外去遮挡快速重建算法[J]. 微电子学与计算机,2024,41(5):31-39. doi: 10.19304/J.ISSN1000-7180.2023.0290
HOU L P,CHEN Y H. A fast reconstruction algorithm for outdoor disocclusion based on neural radiance field[J]. Microelectronics & Computer,2024,41(5):31-39. doi: 10.19304/J.ISSN1000-7180.2023.0290
Citation: HOU L P,CHEN Y H. A fast reconstruction algorithm for outdoor disocclusion based on neural radiance field[J]. Microelectronics & Computer,2024,41(5):31-39. doi: 10.19304/J.ISSN1000-7180.2023.0290

基于神经辐射场的室外去遮挡快速重建算法

A fast reconstruction algorithm for outdoor disocclusion based on neural radiance field

  • 摘要: 三维重建和渲染是虚拟现实的关键技术,如何提高三维重建的质量并降低三维重建的成本是虚拟现实的重要问题。神经辐射场可以在保证高质量重建的同时降低重建的门槛,使得用户仅仅通过手机相机等移动设备采集多视角图像,就可以重建出照片级真实度的物体和场景。然而神经辐射场的重建效果依赖于实验室理想环境中采集的数据。在真实室外环境中,重建场景中会出现暂态物体遮挡,并且用户对重建的速度有较高的要求,这两点神经辐射场均不能很好地处理。针对这两点局限,提出了一种神经辐射场的改进模型FoRF,结合多分辨率哈希编码并修改网络结构以加快建模速度,引入暂态嵌入以处理神经辐射场无法处理暂态物体遮挡的问题。分别在开源数据集、手机拍摄的数据集上进行了实验。实验结果证明了本文模型能够使用移动设备采集的图像重建场景,并具有快速建模和消除暂态物体遮挡对重建结果影响的能力。

     

    Abstract: The field of virtual reality relies on the technologies of 3D reconstruction and rendering. A key objective in this field is to enhance the quality of 3D reconstruction while also reducing the costs associated with this process. The neural radiance field technique has emerged as a potential solution to this problem, as it can generate high-quality reconstructions with reduced reconstruction thresholds by using multi-view images captured from mobile devices like cameras. However, the success of the neural radiation field approach is largely dependent on laboratory conditions and can encounter difficulties when used in real outdoor environments due to transient object occlusion. Furthermore, existing neural radiation fields fail to meet users' expectations for fast reconstructions. To overcome these challenges, this study introduces a new model named FoRF, which incorporates multi-resolution hash encoding and modifies the network structure to accelerate the modeling speed. To address the issue of transient object occlusion, transient embedding is introduced. The experimental results demonstrate that the proposed model effectively reconstructs scenes from images captured using mobile devices and can rapidly model and eliminate the impact of transient object occlusion on reconstruction results.

     

/

返回文章
返回