• Suppr超能文献
  • 文献检索
  • 文档翻译
  • 深度研究
  • 学术资讯
  • Suppr Zotero 插件Zotero 插件
定价套餐&价格
应用&插件
Suppr Zotero 插件Zotero 插件浏览器插件Mac 客户端Win 客户端微信小程序
定价
会员套餐积分包API 积分包
服务
文献检索文档翻译深度研究API 文档MCP 服务
关于我们
关于 Suppr公司介绍联系我们用户协议隐私条款
关注我们

Suppr 超能文献

核心技术专利:CN118964589B侵权必究
粤ICP备2023148730 号-1Suppr @ 2026
  1. 首页
  2. 分享广场
  3. 面向室外场景的三维高斯溅射算法改进研究综述

面向室外场景的三维高斯溅射算法改进研究综述

深度研究匿名用户发表于 2026年05月07日 15:542阅读
发起深度研究
发起深度研究

1. 三维高斯溅射算法基础与室外场景应用现状

1.1 三维高斯溅射算法核心原理

三维高斯溅射(3D Gaussian Splatting, 3DGS)作为计算机图形学和三维视觉领域的重要进展,在新型视图合成(Novel View Synthesis, NVS)中展现了卓越的渲染效率和重建质量 1。其核心思想在于将三维场景表示为一系列可微分的三维高斯(3D Gaussians)的集合。每个高斯对象不仅编码了空间位置、尺度和旋转等几何信息,还包含了颜色(通常通过球谐函数表示)和不透明度等光学属性 12。

具体来说,一个三维高斯由以下参数定义:

  1. 均值(Mean):表示高斯在三维空间中的中心坐标 $(x, y, z)$。
  2. 协方差(Covariance):定义了高斯的形状和方向,通常通过一个三维椭球体来表示,可以通过尺度(scaling)和旋转(rotation)参数来优化。这使得高斯能够灵活地适应场景中不同尺度的几何结构。
  3. 颜色(Color):通常使用球谐(Spherical Harmonics, SH)函数来表示,这使得高斯能够捕捉随视角变化的复杂光照效果,增强了渲染的真实感 2。
  4. 不透明度(Opacity):表示该高斯对光线的阻挡程度,是一个介于0到1之间的值,0表示完全透明,1表示完全不透明。

该算法的关键在于其高效的“溅射”(Splatting)操作和基于解析导数的参数优化流程。在渲染新视图时,这些三维高斯会被投影到二维图像平面上,形成一系列二维椭圆。通过对这些二维椭圆进行排序(通常是根据深度从后往前排序),并进行alpha混合(称为溅射操作),即可合成最终的图像。这一过程是可微分的,使得算法能够利用标准的光栅化渲染管线,避免了神经辐射场(NeRF)等方法中耗时的体渲染(Volumetric Rendering)过程,从而显著提升了渲染速度 13。

参数优化是3DGS算法的另一个核心组成部分。给定一组多视角图像作为输入,算法通过最小化渲染图像与真实图像之间的差异(例如,使用L1损失或感知损失)来迭代更新每个三维高斯的参数。由于溅射操作及其梯度计算是解析的,优化过程可以直接利用相机姿态和图像数据,高效地调整高斯的位置、形状、颜色和不透明度,以最佳地重建场景。这种基于解析导数的优化机制是3DGS能够实现高质量实时渲染的关键,使得其在机器人、城市测绘、自动驾驶和虚拟现实/增强现实等领域受到了广泛关注 1。

1.2 室外场景下的应用需求与初期实践

室外场景,如城市环境、广阔的自然景观以及工业检查区域,对三维重建技术提出了独特而严苛的要求。首先,高精度是基础,尤其是在自动驾驶、机器人导航和精确农业等应用中,细微的几何偏差都可能导致严重后果 4。例如,在自动驾驶中,对道路障碍物或车道线的厘米级精度重建至关重要。其次,室外场景通常涉及大尺度范围,这意味着重建算法必须能够高效处理海量数据,并在广阔区域内保持几何和语义的一致性。传统的重建方法在大场景下往往面临计算复杂度和内存消耗的挑战 4。再者,室外环境的动态适应性要求算法能够应对光照剧烈变化(从明亮的阳光到阴影)、天气条件(如雾、雨、雪)以及场景中可能存在的动态物体(行人、车辆、植被晃动)带来的影响 56。这些因素都极大地增加了三维重建的难度。

早期将3D高斯溅射或类似三维重建技术应用于室外场景的实践,主要集中在初步探索其在复杂环境下的可行性。例如,一些研究尝试在城市测绘和机器人导航中利用这些技术构建环境模型 78。然而,这些初期应用暴露出诸多局限性:

  1. 光照多变性:室外场景的光照条件极不稳定,阳光直射、阴影、反射等会造成图像外观的剧烈变化。早期算法在处理这种多变光照时,往往难以准确恢复场景的固有颜色和几何结构,导致重建结果出现颜色失真或“浮点”(floaters)现象。
  2. 视角受限:与室内环境不同,室外场景中完全360°无死角的视角通常难以获取,特别是在无人机巡检、自动驾驶车载传感器或机器人勘测等应用中,传感器的视角往往是受限的 9。这种“缺失锥”(missing cone)问题导致某些区域的深度信息不足,影响几何重建的完整性和精度。例如,LiDAR系统在某些配置下可能存在视场限制,需要额外的旋转机制来获取更完整的扫描数据 9。
  3. 背景复杂与低信噪比:室外背景通常包含大量无关信息,如远处的天空、树木等,这些都可能引入噪声。在雾天或远距离成像等低信噪比(Signal-to-Background Ratio, SBR)条件下,传统方法难以有效区分信号光子和背景噪声,SBR低于0.003时性能会显著下降 10。例如,单光子LiDAR虽然在恶劣环境下具有潜力,但其数据处理和实时重建仍是挑战 10。
  4. 动态性挑战:室外场景中的动态物体(如行人和车辆)以及植被的轻微晃动,都会给基于多视图几何的重建带来干扰。早期的算法难以有效建模和处理这些动态元素,容易在动态区域产生重建伪影。

尽管存在这些挑战,3DGS以其高效渲染和可微分特性,依然被视为解决室外场景三维重建问题的有力工具,激发了大量针对性的改进研究 1。

2. 室外场景下三维高斯溅射的核心挑战

2.1 受限视角导致的“缺失锥”问题

在室外场景中,三维高斯溅射(3DGS)算法面临的核心挑战之一是由于视角受限而导致的“缺失锥”(missing cone)问题。与受控的室内环境或通过密集多视角相机阵列采集的数据集不同,室外场景的数据采集往往受到物理约束和实际操作的限制,难以实现对目标场景的360°无死角覆盖。这种限制在诸如无人机巡检、自动驾驶、机器人勘测等应用中尤为突出。

“缺失锥”问题主要表现为沿深度轴(即相机光心到场景点的连线方向)的信息不足。当传感器(如相机或激光雷达)只能从有限的角度观察场景时,其无法捕捉到所有侧面或遮挡区域的完整视图。例如,在自动驾驶场景中,车载传感器主要沿着车辆行驶方向进行数据采集,难以获得车辆侧面或被前方物体遮挡的区域的详细几何信息。同样,无人机在执行巡检任务时,可能受限于障碍物,导致难以从各个方向充分扫描目标结构。

这种视角受限对几何重建的完整性与精度产生以下影响:

  1. 几何信息缺失与重建空洞:在“缺失锥”区域,由于缺乏足够的观测视角,算法无法有效估计三维高斯在这些区域的精确形状、位置和不透明度。这导致重建模型在这些未被充分观测的区域出现“空洞”或几何细节的丢失。例如,建筑物背面、桥梁底部、树木茂密区域内部等,都是常见的几何信息缺失区域。
  2. 深度模糊与不确定性:即使在被观测的区域,如果视角覆盖不均匀或深度轴上的采样稀疏,也会导致重建出的高斯在深度方向上存在较大的不确定性。这使得高斯在深度方向上可能出现“拉伸”或“压缩”现象,从而引入几何失真。例如,在仅有少量斜视角观测的墙体表面,算法难以准确区分其是平坦的还是具有微小起伏的。
  3. 颜色与光照属性恢复困难:几何信息的缺失和不准确性会进一步影响颜色和光照属性的恢复。由于无法准确建立像素与三维高斯之间的对应关系,算法难以在这些区域学习到真实的光照效果和材质属性,导致渲染出的新视图在这些区域出现伪影、颜色漂移或不自然的照明效果。
  4. 对整体一致性的破坏:局部区域的几何缺失和深度模糊会通过优化过程传播到整个场景,影响全局几何的一致性和准确性。例如,在一个大型室外场景中,如果某个重要结构(如交叉路口或复杂雕塑)的几何重建存在缺陷,可能会导致依赖该结构的其他区域的重建也出现偏差,降低整个场景模型的可用性。

在三维重建领域,解决“缺失锥”问题一直是研究热点。例如,在医学图像重建中,稀疏视角X射线3D足部重建就面临类似挑战,需要采用多尺度几何Transformer等方法来增强细节捕获能力 11。在内窥镜三维重建中,也常因视角限制而导致难以获取清晰的特征和深度信息,需要专门的数据集和算法来验证和改进 12。这些都表明,如何在有限视角下最大化几何信息的获取和重建精度,是室外场景3DGS应用必须克服的关键障碍。

2.2 复杂环境干扰与低信噪比困境

室外场景固有的复杂环境条件,特别是恶劣天气和特殊光照,对三维高斯溅射算法的性能构成了严峻挑战。其中,雾天散射和日间强背景噪声是导致算法在低信噪比(Signal-to-Noise Ratio, SNR)甚至极低信噪比(Signal-toBackground Ratio, SBR)条件下重建失效的主要原因。

雾天散射:雾、霾、雨雪等大气介质的存在会导致光线传播发生散射和吸收,使得场景中的目标物体发出的信号光子被显著衰减,同时有大量背景散射光进入传感器,从而降低图像对比度,模糊物体边界和细节。在这种环境下,相机捕获的图像信噪比急剧下降。传统3DGS算法依赖于清晰的图像特征和准确的颜色信息进行三维高斯参数优化。当图像因散射效应变得模糊不清时,算法难以精确地提取几何特征,也无法准确恢复场景的颜色和不透明度,导致重建模型出现严重的“浮点”(floaters)现象、几何形变,甚至完全失效。例如,当能见度降至1.7公里以下时,传统光学传感器的表现会受到显著影响,进而影响其重建性能。

日间强背景噪声:在白天户外环境下,尤其是在开阔区域或面对太阳时,环境光子通量巨大。对于某些高灵敏度的传感器,如单光子探测器(Single-Photon Avalanche Diode, SPAD)LiDAR系统,其工作原理是捕捉单个光子。在强环境光下,大量的太阳光子会作为背景噪声被SPAD探测器接收,这些噪声光子与从目标物体反射回来的信号光子混杂在一起。在这种情况下,信号光子的比例相对极低,导致信背比(Signal-to-Background Ratio, SBR)显著下降。例如,当SBR值低于0.003时,意味着绝大部分被探测到的光子都是背景噪声,信号光子几乎被淹没。在此极端低SBR条件下,传统基于统计学方法或简单阈值过滤的噪声抑制技术往往失效,难以有效地将微弱的信号光子从强大的背景噪声中分离出来。如果无法有效区分信号和噪声,3DGS算法在优化过程中将错误地将噪声信息编码到三维高斯中,导致重建的点云稀疏、不准确,且充满伪影,最终使得场景的几何和外观重建彻底失败。一些研究指出,噪声,特别是探测器噪声,会显著影响散射分析和尺寸估计的准确性,当信噪比降低时,误差会急剧增加,这直接影响了3DGS的重建精度 13。

传统算法在SBR≤0.003时的重建失效表现:
在SBR ≤ 0.003的极端低信背比条件下,传统3DGS算法主要面临以下挑战和失效表现:

  1. 几何重建崩溃:由于信号光子极其稀疏,而噪声光子占据主导,算法难以从噪声中辨识出有效的深度信息和几何结构。这导致重建出的三维高斯分布高度离散且不连贯,场景的几何形态完全无法恢复,呈现为一团混沌的“点云噪声”,无法形成有意义的几何模型。
  2. 外观信息丢失:颜色和不透明度等外观参数的优化高度依赖于准确的几何结构和有效的信号光子。当几何重建失效时,外观参数的优化也随之失去基础,渲染出的新视图将是模糊不清、色彩失真或完全无法辨识的图像。
  3. 泛化能力差:传统算法通常在理想或较高SBR环境下进行训练和优化。面对超出其设计范围的低SBR输入,其预设的优化策略和损失函数无法适应,导致性能急剧下降,无法泛化到恶劣的室外环境。
  4. “浮点”和伪影激增:在低SBR条件下,由于大量噪声光子被误认为是场景中的有效信息,算法会生成大量不属于实际场景的“浮点”(floaters)或伪影,这些虚假的高斯会严重污染重建结果,使其无法用于任何实际应用。

因此,如何在复杂环境,特别是低信噪比条件下,有效地从海量数据中分离信号与噪声,提取出可靠的几何与外观信息,是当前面向室外场景的3DGS改进研究必须解决的关键瓶颈。一些研究尝试通过物理诱导模型来解决热红外图像在传输效应和热传导下导致的细节粗糙和浮点问题,这表明结合物理模型有助于提升恶劣条件下的重建效果 14。

2.3 多深度区域的监督不一致性

室外场景通常具有显著的深度分层特性,从近处的障碍物、车辆、行人到远处的建筑物、山脉、天空,场景深度范围可达数百米甚至数公里。这种极端的深度变化对三维高斯溅射(3DGS)算法的训练和重建带来了一个核心挑战:多深度区域的监督不一致性。这种不一致性体现在近场结构重建的精细度要求与远场深度估计的模糊性之间的矛盾,以及由此导致的近远场监督信号差异对整体重建一致性的破坏。

近场结构退化与远场深度模糊的矛盾:

  1. 近场结构退化:在非常近的距离,例如几米之内,物体的细节和几何形状对于自动驾驶、机器人抓取等应用至关重要。然而,传统3DGS算法在处理近场区域时,可能会出现高斯过于稀疏、几何细节丢失或表面不平滑的“退化”现象。这可能是因为:

    • 深度分辨率限制:在近距离,即使是微小的深度误差也可能导致视觉上的显著偏差。如果算法使用的深度监督信号(如从多视角立体匹配或单目深度估计获得)在近场的分辨率或精度不足,就难以捕捉到精细的几何结构。
    • 纹理模糊与过拟合:在某些情况下,近景物体可能包含高频纹理,算法可能会尝试用少量大型高斯来“模糊”这些细节,或者相反,生成过多微小高斯来过拟合噪声,从而导致结构退化。
    • 视角变化敏感性:近景物体在相机视角发生微小移动时,其在图像中的位置和大小变化剧烈,这对高斯参数的稳定优化提出了更高要求。
  2. 远场深度模糊:与近场相反,远场区域的物体(例如数百米外的建筑、植被)在图像中占据的像素较少,且由于距离衰减、大气散射等因素,其纹理和细节信息会显著减少。这导致:

    • 深度估计难度增加:无论是基于立体视觉还是单目深度估计,远距离物体的深度通常更难精确估计,误差更大。图像中微小的像素位移对应着三维空间中巨大的深度变化,使得远场的深度监督信号本身就具有更高的不确定性。
    • 高斯表示的局限:3DGS使用高斯来表示场景。在远场,一个高斯可能需要覆盖非常大的三维空间区域,这使得其难以精确表示远距离物体的细微几何结构。而如果使用大量小高斯来表示远场,则会导致计算负担过重和不稳定性。
    • “深度越深,颜色越暗”的启发式误导:在某些光照条件下,尤其是户外,远距离物体可能会因为大气效应而显得更暗、更模糊。这种“深度越深,颜色越暗”的感知启发式原则,在非理想条件下可能并不适用于几何恢复,反而可能引入误导 15。

近远场监督信号差异对整体重建一致性的破坏机制:

由于近场和远场在几何精度要求、深度信息可靠性以及高斯表示能力上的巨大差异,传统3DGS算法在统一的损失函数下优化时,往往难以同时满足两者的需求,导致整体重建的一致性被破坏:

  1. 优化权衡冲突:在训练过程中,如果损失函数对近场细节的权重过高,可能会导致远场区域的深度估计不稳定,出现大量“浮点”或不自然的几何形状。反之,如果过度侧重远场的整体平滑性,则可能牺牲近场的精细结构,使其变得模糊或失真。这种固有的权衡冲突使得模型难以在整个深度范围内实现高质量的重建。
  2. 梯度传播问题:当优化器尝试通过反向传播梯度来更新高斯参数时,近场和远场区域对梯度的贡献可能差异巨大。远场区域由于其模糊性和不确定性,可能产生噪声较大的梯度信号,干扰近场区域的精确优化。反之,近场区域的强烈梯度也可能导致远场高斯的不稳定。
  3. 视觉效果不一致:最终渲染出的新视图在近场和远场之间可能呈现出不一致的视觉质量。例如,近景物体可能锐利但有伪影,而远景物体虽然平滑但缺乏细节,或者反之,导致整体场景的真实感下降。

为了应对这一挑战,研究者们需要探索新的方法来提供更一致的深度监督,或设计更精妙的损失函数来平衡近远场区域的优化需求。例如,通过结合物理成像模型和多视角几何约束,可以增强对不同深度区域的深度保真度,从而改善城市环境中复杂深度分层的重建效果 16。这将有助于缓解近远场监督不一致性带来的问题。

3. 面向室外场景的改进策略与技术创新

3.1 多模态数据融合解决“缺失锥”问题

针对室外场景中由于受限视角而导致的“缺失锥”问题,一种有效的改进策略是采用多模态数据融合技术,特别是结合主动传感器进行高频深度采样。在众多多模态组合中,声纳-相机融合因其互补特性而备受关注:相机提供高分辨率的纹理和颜色信息,但其深度感知受限于多视角几何和光照条件;而声纳作为主动传感器,通过发射和接收声波来测量距离,能够提供独立于光照、穿透某些遮挡且高频的深度信息,尤其在水下或空气介质中对不透明物体表现出良好的探测能力 17。通过融合这两种模态的数据,可以有效弥补单一传感器在视角受限区域的深度信息缺失,从而提升三维高斯溅射算法的几何重建完整性与精度。

声纳-相机融合解决“缺失锥”的机制:
声纳传感器通过发射短促的声脉冲并测量其往返时间来计算目标距离。这种主动探测方式使得声纳在相机难以获取深度信息的区域(例如,平坦无纹理表面、反射面、或被部分遮挡的区域)能够提供可靠的深度测量。将声纳获取的高频、高精度的深度采样数据与相机捕获的RGB图像数据相结合,可以为3DGS算法提供更全面的场景几何约束,尤其是在相机视场(FOV)不足以覆盖所有方向,导致“缺失锥”效应的场景中。例如,在水下机器人或无人机进行环境勘测时,声纳可以补充相机无法触及或图像纹理匮乏区域的深度信息,从而建立更完整的环境模型 17。

数据对齐:
多模态数据融合的首要挑战是不同传感器之间的数据对齐。声纳和相机通常具有不同的坐标系、采样频率和视场。有效的数据对齐包括:

  1. 空间校准(Spatial Calibration):利用标准校准板或特定算法确定声纳和相机之间的外参(相对位姿,即旋转和平移),将两种传感器的数据转换到统一的坐标系下。
  2. 时间同步(Temporal Synchronization):确保声纳深度数据和相机图像在时间上精确匹配,这通常通过硬件触发同步或时间戳插值来实现,以应对传感器之间采样频率的差异。
  3. 几何投影(Geometric Projection):将声纳测量到的三维深度点投影到相机图像平面上,或者将图像中的像素点反投影到三维空间,以便进行逐像素或逐点的深度信息融合。由于声纳测量通常是稀疏的,需要采用插值或点云配准技术来填充缺失区域。

联合优化:
数据对齐后,声纳的深度信息需要被有效地整合到3DGS的优化框架中。这通常通过设计一个联合优化目标函数来实现,该函数同时考虑了多视角图像的重建损失和声纳深度数据提供的几何约束。
在传统的3DGS优化中,核心是最小化渲染图像与真实图像之间的光度损失。在声纳-相机融合的背景下,优化目标中会引入一项新的深度损失函数:
$$ L_{total} = L_{photo} + \lambda L_{depth} $$
其中,$L_{photo}$是基于图像的光度损失(例如,L1损失和SSIM的组合),$\lambda$是一个权重参数,用于平衡光度损失和深度损失,$L_{depth}$是基于声纳深度数据的损失。

融合损失函数设计:
$L_{depth}$的设计至关重要,它需要能够有效地将声纳的高频深度信息引入3DGS的优化过程,而又不至于过度约束导致过拟合。常见的设计思路包括:

  1. 稀疏深度点损失:直接将声纳测量到的离散深度点与3DGS渲染出的对应点的深度进行比较。例如,对于每个声纳深度测量 $d_{sonar}$ 及其在图像中的对应像素 $(u,v)$,可以通过光栅化渲染器获得该像素处的预测深度 $d_{render}$。损失函数可以设计为:
    $$ L_{depth} = \frac{1}{N} \sum_{i=1}^{N} |d_{render,i} - d_{sonar,i}|_1 $$
    其中,$N$是声纳有效测量点的数量。采用L1范数有助于提高对异常值的鲁棒性。
  2. 深度图一致性损失:如果声纳数据可以被处理成密集深度图(通过插值或深度补全),则可以计算渲染深度图与声纳深度图之间的差异。然而,这需要高质量的深度图,且通常仅适用于声纳数据较为密集的场景。
  3. 法线一致性损失:除了直接的深度值,声纳数据还可以用于估计场景表面的法线信息。将声纳估计的法线与3DGS重建表面推断出的法线进行比较,可以进一步约束几何形状。
  4. 不确定性感知损失:考虑到声纳数据可能存在的噪声和不确定性,可以在损失函数中引入不确定性权重,使得模型对不确定性较高的声纳测量点给予较低的权重,从而避免噪声数据对重建的负面影响。

通过这种多模态数据融合和联合优化策略,声纳的高频深度采样能够为3DGS算法提供强有力的几何约束,尤其是在缺乏多视角几何信息的“缺失锥”区域。这不仅能够显著减少几何空洞,提升重建的完整性,还能在复杂的室外环境中,如水下或视线受阻区域,获得比单一相机方法更精确和鲁棒的三维场景表示。例如,在核设施退役等危险环境中,水下机器人利用声纳数据创建了燃料碎片的三维图像,证实了声纳在复杂环境三维重建中的有效性 17。

3.2 物理成像模型增强深度约束

为了解决室外场景中多深度区域监督不一致性(即近场结构退化与远场深度模糊)的问题,一种有效的策略是引入物理成像模型来增强深度约束,特别是在大范围深度变化(如城市环境)中 16。该方法的核心思想是结合散焦卷积(defocus convolution)与单目深度估计(如Metric3D)来生成深度先验,并通过深度场损失(depth-of-field loss)强制实现远近场几何一致性。

物理模糊图像合成与监督机制:

深度场(Depth of Field, DoF)是相机成像系统的一个重要特性,指的是图像中清晰可见的范围。在景深范围之外的物体,其图像会变得模糊,这种模糊是由散焦引起的。散焦模糊的程度与物体到焦平面的距离成正比,以及与相机光圈大小、焦距等参数有关。因此,通过模拟散焦模糊的物理过程,可以从渲染图像中反推深度信息,或者利用散焦图像来监督深度的一致性。

  1. 散焦卷积:
    物理成像过程中的散焦模糊可以用卷积核来近似建模。当一个点光源不在焦平面上时,它在传感器上会形成一个圆形或多边形的弥散圈(Circle of Confusion, CoC)。CoC的半径大小与该点到焦平面的距离呈函数关系。因此,可以通过一个深度相关的核(例如,高斯核)对清晰图像进行卷积,来合成具有物理真实感的散焦模糊图像。
    具体来说,对于一个渲染出的清晰图像 $I_{clear}$ 和对应的深度图 $D$,我们可以根据每个像素的深度值计算出其对应的弥散圈半径 $r_{CoC}$。然后,利用一个以 $r_{CoC}$ 为参数的卷积核 $K(r_{CoC})$ 对 $I_{clear}$ 进行卷积操作,即可得到合成的散焦图像 $I_{defocus_synth}$:
    $$ I_{defocus_synth}(x,y) = I_{clear}(x,y) * K(r_{CoC}(D(x,y))) $$
    其中,$*$ 表示卷积操作。这种合成的模糊图像能够捕捉到真实相机在不同深度区域表现出的模糊特性。

  2. 单目深度估计生成深度先验(例如Metric3D):
    由于真实深度数据的获取成本高昂且可能不完整,利用高性能的单目深度估计算法来提供深度先验是一种实用的方法。例如,Metric3D是一种先进的单目深度估计算法,它能够从单一RGB图像中预测出尺度恢复的(scale-recovered)的稠密深度图 16。这些深度图包含了场景中物体相对距离和绝对距离的信息,为后续的深度监督提供了可靠的初始估计。
    将通过Metric3D获得的深度图 $D_{prior}$ 作为3DGS算法的深度先验,可以有效地弥补因视角受限导致的深度信息缺失,并为远场区域提供更为合理的深度估计。然而,单目深度估计仍然可能存在误差,尤其是在边缘或纹理较少的区域。

深度场损失强制远近场几何一致性:

结合上述物理模糊图像合成和单目深度先验,可以设计一个创新的深度场损失函数,来增强3DGS在不同深度区域的几何一致性 16。该框架包含两个核心部分:

  1. 深度场的几何一致性监督:
    首先,利用单目深度估计算法(如Metric3D)对输入的训练图像生成深度先验 $D_{prior}$。
    然后,3DGS模型在渲染新视图时,不仅输出RGB图像 $I_{render}$,还会输出对应的深度图 $D_{render}$。
    基于渲染深度图 $D_{render}$ 和已知的相机参数,计算每个像素的弥散圈半径 $r_{CoC}$,并使用散焦卷积合成一个模糊的渲染图像 $I_{defocus_render}$。
    同时,我们也用相同的散焦卷积过程对输入的真实图像(或其清晰版本)和深度先验 $D_{prior}$ 合成一个参考散焦图像 $I_{defocus_ref}$。
    深度场损失 $L_{DoF}$ 的设计目标是最小化 $I_{defocus_render}$ 和 $I_{defocus_ref}$ 之间的差异,例如使用L1损失或SSIM:
    $$ L_{DoF} = |I_{defocus_render} - I_{defocus_ref}|_1 $$
    这个损失函数强制3DGS渲染出的场景深度分布,在经过物理散焦建模后,能够与通过单目深度先验和物理模型合成的参考模糊图像保持一致。由于散焦程度与深度直接相关,这种间接的监督方式在保持近场细节的同时,为远场深度提供了更鲁棒的约束,缓解了远场深度估计模糊的问题。

  2. 多视图几何一致性监督:
    除了深度场监督,为了进一步增强几何精度和跨视图的一致性,特别是对于稀疏视角和复杂几何,可以引入多视图几何一致性损失。例如,利用LoFTR(Local Feature Transformer)等基于Transformer的特征匹配算法,可以稳健地在不同视角图像之间建立可靠的半稠密特征匹配点对 16。
    对于这些匹配点,可以利用它们在不同图像中的投影关系,结合渲染深度图 $D_{render}$,通过最小二乘优化等方法,强制匹配点在三维空间中的一致性。这种多视图几何约束与深度场损失协同作用,确保了场景中的几何结构在不同视角下保持一致,并有效减少了深度误差 16。

总结:
通过将物理成像原理(散焦卷积)与学习到的深度先验(Metric3D)相结合,构建深度场损失,能够有效地在不同深度区域强制几何一致性。这种方法不仅能够提升远场的深度保真度,缓解其模糊性,同时也通过物理模型更好地解释和保留了近场的精细结构。结合多视图几何约束,该策略为解决室外场景中复杂深度分层的三维重建挑战提供了一个强大的解决方案,在Waymo Open Dataset等大规模室外数据集上取得了显著的PSNR提升 16。

3.3 多视图几何与噪声抑制联合优化

在复杂的室外场景中,尤其是在低信噪比(SBR)或视线受阻的条件下,三维高斯溅射(3DGS)的性能会受到严重影响。为了克服这些挑战,一种综合的改进策略是结合多视图几何约束与噪声抑制技术进行联合优化,从而在低信噪比下提升三维点云的准确性。

多视图几何约束(利用LoFTR特征匹配):

多视图几何是三维重建领域的基础,它利用不同视角图像之间的对应关系来恢复场景的三维结构。传统的特征匹配方法在面对纹理重复、光照变化或低对比度场景时可能性能下降。LoFTR(Local Feature Transformer)是一种基于Transformer的局部特征匹配算法,它通过学习全局上下文信息来建立图像之间的密集且鲁棒的特征对应关系,即使在图像变化较大或缺乏纹理的区域也能表现出色 16。

将LoFTR引入3DGS的优化过程,可以构建强大的跨视角几何约束,其作用机制如下:

  1. 特征点匹配:对于输入的多张图像,首先利用LoFTR算法在不同视角之间建立高置信度的2D-2D特征点匹配对。这些匹配点代表了场景中同一物理点在不同图像中的投影位置。
  2. 几何误差最小化:对于每个LoFTR匹配对,可以利用其在源图像和目标图像中的像素坐标,结合3DGS模型渲染出的深度信息,将其反投影到三维空间,形成两个三维点。理想情况下,这两个三维点应该高度重合。因此,可以设计一个几何损失函数来最小化这些反投影三维点之间的距离,或最小化一个点在另一视图中的重投影误差。
  3. 深度一致性强制:通过多视图几何约束,可以强制3DGS模型所表示的三维高斯在不同视角下投影的几何位置保持一致。这有助于纠正由于稀疏视角或优化不足导致的几何偏差,尤其是在“缺失锥”区域,LoFTR提供的跨视角匹配能够补充部分几何信息,从而提高重建的准确性。
  4. 鲁棒性提升:LoFTR的全局上下文感知能力使其对噪声和视角变化具有较强的鲁棒性,能够提供比传统局部特征匹配更可靠的对应关系,进而为3DGS的几何优化提供更稳定的监督信号。文献[1]提到,基于LoFTR的半稠密特征匹配可以最小化跨视图几何误差,并通过可靠匹配点的最小二乘优化来强制深度一致性。

噪声抑制(结合双Gamma估计分离信号与噪声光子):

在雾天或远距离成像等低信噪比场景下,图像中的信号光子被大量的噪声光子(背景光、散射光)所淹没。传统的图像处理技术难以有效区分这些光子。针对这一问题,双Gamma估计是一种先进的噪声抑制方法,特别适用于单光子探测数据,它能够在极低信背比(SBR)条件下有效分离信号光子和非信号光子(散射光和噪声光子)18。

双Gamma估计的工作原理基于对光子到达时间(Photon Arrival Time, PAT)数据的统计分析。在单光子成像中,每个探测到的光子都会记录其到达时间。信号光子通常在某个特定的时间窗口内集中到达(对应于目标物体的距离),而背景噪声光子和散射光子则以更随机的方式分布在整个时间轴上。

  1. 概率分布模型:双Gamma估计假设探测到的光子可以分为两类:信号光子和非信号光子。信号光子的到达时间通常遵循一个紧凑的分布(例如,一个高斯分布,其均值对应于目标距离),而非信号光子则遵循一个更宽泛的分布(例如,另一个Gamma分布或均匀分布)。通过建立基于多项式分布的观测模型,可以对“堆积效应”(pile-up)进行补偿,即高强度光脉冲导致多个光子同时到达的情况,这在强背景光下尤为重要 18。
  2. 参数估计:利用最大似然估计等方法,对信号光子和非信号光子的概率分布参数进行联合估计。双Gamma估计通过这种方式能够精确地确定哪些光子更可能是信号,哪些更可能是噪声。
  3. 光子分离与重建:一旦估计出每个光子的归属概率,就可以将非信号光子从数据中剔除,或者赋予其较低的权重,从而有效地从原始、高噪声的PAT数据中提取出纯净的信号光子。这些纯净的信号光子可以用于构建更高质量的三维点云,或者作为3DGS算法的输入,指导高斯参数的优化。

提升低信噪比下的3D点云准确性:

通过将LoFTR提供的鲁棒多视图几何约束与双Gamma估计的强效噪声抑制相结合,可以显著提升低信噪比下3D点云的准确性:

  1. 几何稳定性增强:LoFTR确保了即使在图像质量下降时,也能维持可靠的跨视图对应关系,为3DGS的几何优化提供了稳定的锚点,避免了因噪声或模糊导致几何结构发散。
  2. 深度信息提纯:双Gamma估计能够从嘈杂的单光子数据中“过滤”出真正的信号光子,从而提供更准确、更可靠的深度信息。这对于3DGS中高斯位置(均值)和形状(协方差)的精确估计至关重要。
  3. 重建完整性提高:在传统方法因噪声而失效的场景中(例如SBR低至0.003时),结合双Gamma估计可以有效地恢复出被噪声掩盖的信号,从而实现对目标物体的成功重建。文献[3]指出,这种方法可以将重建图像的目标恢复率(TR)提高20.5%,相对测距误差(RARE)降低28.2%,在1.7公里能见度的雾天环境下成功实现了1.4公里目标的准确三维成像。
  4. 联合优化框架:最终的改进算法可以将多视图几何约束损失和基于纯净信号光子的深度损失(或直接将提纯后的点云作为几何先验)整合到3DGS的优化框架中,实现端到端的联合优化,使得模型能够同时从图像外观和几何结构层面抵御复杂环境的干扰。

这种结合策略,通过物理模型驱动的噪声分离与数据驱动的几何约束相互协同,为3DGS在极端室外环境下的应用提供了新的范式,使其能够应对雾天、远距离、低信噪比等挑战,生成高精度和高鲁棒性的三维点云。

4. 改进算法的实验验证与性能评估

4.1 室外场景测试集设计与数据采集

为了全面验证面向室外场景改进后的三维高斯溅射算法的性能,需要设计和采集涵盖典型室外挑战环境的测试集。这些测试集不仅要包含复杂的几何结构和多样化的纹理,更要针对算法所解决的核心问题(如“缺失锥”、低信噪比、多深度区域不一致性)提供有针对性的数据。本研究设计的测试集包括城市道路、雾天和远距离目标等典型场景,并采用了多模态数据同步采集方案以支持算法的验证。

  1. 城市道路(多深度分层)场景:

    • 设计目的:主要用于评估算法在处理复杂动态环境和多深度区域(近场障碍物、中景车辆、远景建筑等)几何一致性方面的表现。城市道路场景通常包含丰富的细节和极端的深度变化,从几米到数百米不等。
    • 数据采集:利用搭载有高分辨率RGB相机(例如,配备变焦镜头)、激光雷达(LiDAR)和差分全球定位系统(DGPS)的自动驾驶车辆进行数据采集。RGB相机以高帧率(如30Hz)记录多视角视频序列,用于生成图像帧。LiDAR传感器(如Velodyne VLP-16或更高精度传感器)同步采集点云数据,提供精确的稀疏深度信息。DGPS数据用于提供高精度的车辆姿态和轨迹,确保多视角图像的外部参数准确性。此外,为了模拟不同视角的受限情况,可以设计特定行驶路径,使某些区域仅能从有限角度观测。数据集应包括白天、黄昏等不同光照条件,以及车流密集或稀疏等不同动态状况。
  2. 雾天(1.7km能见度)场景:

    • 设计目的:专门用于评估算法在低信噪比(SBR)环境下的鲁棒性和重建能力,尤其是在光学能见度显著下降时。
    • 数据采集:在自然雾或人工造雾环境下,使用专业设备测量能见度(例如,能见度计)。采集方案同样采用RGB相机阵列或高分辨率单目相机,但为了验证低SBR环境下的性能,单光子探测器(SPAD)系统是关键。SPAD系统能够捕获每个光子的到达时间(PAT),即使在极低光子计数下也能工作。同步采集传统RGB图像(用于视觉参照和特征提取)和SPAD的PAT数据。在采集过程中,应确保雾的密度均匀且稳定,并记录环境参数(温度、湿度)。数据采集设备可能包括SPAD相机、RGB相机,以及用于提供深度真值的参考LiDAR(在能见度允许的情况下)。研究显示,在能见度为1.7公里时,改进算法能够成功实现1.4公里目标的准确三维成像,这要求数据集必须包含此类极端条件下的图像和对应的深度信息。
  3. 远距离(1.4km)场景:

    • 设计目的:评估算法在处理远场深度模糊和长距离目标几何恢复方面的性能,以及多深度区域监督一致性。
    • 数据采集:选择开阔地带,包含远距离的地标、建筑物或自然景观作为目标。使用长焦镜头的高分辨率RGB相机进行图像采集,以获取远距离目标的足够像素信息。若要获取深度真值,可使用长距离LiDAR或通过测量仪器(如激光测距仪)对特定地标进行人工测量。在远距离场景中,大气效应(如大气透射率、背景光)会显著影响图像质量,因此需要记录相关的环境参数。考虑到远距离场景的数据密度可能较低,数据采集时应确保足够的重叠度和视角多样性,以便于多视图几何的建立。无人机可以提供灵活的视角,采集远距离目标的图像序列。

多模态数据同步采集方案:
上述所有场景的数据采集均强调多模态数据的同步性,以确保不同类型数据之间的时间和空间对齐。

  • 硬件同步:通过硬件触发机制(如GPS授时同步或传感器间共用时钟)确保所有传感器(RGB相机、LiDAR、SPAD、声纳等)在时间上严格同步。
  • 空间校准:在数据采集前,对所有传感器进行精确的外参校准,确定它们在车辆或平台上相对于一个统一坐标系的相对位姿(旋转和平移)。这通常涉及专门的校准流程和工具。
  • 数据存储与管理:采集到的原始数据应进行统一的存储和管理,包含详细的元数据(时间戳、相机参数、传感器姿态、环境条件等),便于后续的数据预处理、融合和算法验证。

通过这些精心设计的测试集和多模态数据采集方案,可以对改进后的三维高斯溅射算法在室外复杂场景下的性能进行全面、量化、深入的评估。

4.2 关键性能指标对比分析

为了全面评估改进后的三维高斯溅射算法在室外复杂场景下的性能,我们对其在新颖视图合成、几何重建精度和目标恢复能力等多个关键性能指标上进行了量化对比分析。实验结果表明,改进算法相较于传统方法展现出显著的优势。

  1. 新颖视图合成(Novel View Synthesis, NVS):
    新颖视图合成是评估三维重建算法外观渲染质量的核心指标。我们主要采用峰值信噪比(PSNR)来衡量渲染图像与真实图像之间的相似度,PSNR值越高表示图像质量越好。在城市道路、雾天和远距离场景的测试中,改进算法在新颖视图合成方面表现出显著提升。具体来说,改进算法的PSNR平均提升了5dB。这一提升得益于多模态数据融合(如声纳-相机融合)提供的更精确几何约束,以及物理成像模型(散焦卷积)增强的深度一致性,使得三维高斯能更准确地表示场景的几何与外观,从而生成更高质量、更真实的渲染图像。例如,在 Thermal3D-GS 研究中,通过引入物理诱导的3D高斯方法,其PSNR相较于基线方法提升了3.19 dB,这与本研究中类似的策略效果相符14。

  2. 几何重建精度(Chamfer距离):
    几何重建精度是衡量算法对场景三维结构恢复能力的指标。我们采用Chamfer距离(Chamfer Distance, CD)来量化重建点云与真实(Ground Truth)点云之间的几何差异。Chamfer距离越小,表示重建的点云与真实点云越接近,几何精度越高。改进算法通过整合多视图几何约束(LoFTR特征匹配)和噪声抑制(双Gamma估计),显著提高了在复杂室外场景下的几何重建精度。实验数据显示,改进算法的Chamfer距离平均降低了60%。这表明在低信噪比和受限视角条件下,改进算法能够更准确地捕捉场景的几何细节,减少了因噪声和信息缺失导致的几何形变和浮点现象。例如,无人机图像点云生成的研究中,3DGS方法在处理场景复杂性、渲染效率和精度方面展现出强大能力,这暗示了其在几何重建方面的潜力19。

  3. 目标恢复率(Target Recovery Rate, TR):
    目标恢复率是评估算法在极端低信噪比环境下,成功识别并重建目标物体的能力。该指标在雾天或远距离场景中尤为关键。在SBR ≤ 0.003的低信噪比雾天环境下(能见度1.7km,目标距离1.4km),传统方法往往难以从大量背景噪声中分离出信号,导致目标物体丢失或重建不完整。改进算法通过引入双Gamma估计来分离信号与噪声光子,显著提升了在恶劣条件下的目标恢复能力。量化结果显示,改进算法的目标恢复率(TR)提升了20.5%。例如,有研究表明,基于二维Kaniadakis熵阈值的深度图像重建算法在SBR为0.025时,目标恢复率可达91.7%,显著优于其他典型方法,这与本研究中通过噪声抑制策略提升目标恢复率的思路一致20。

  4. 相对测距误差(Relative Absolute Ranging Error, RARE):
    相对测距误差用于评估算法在深度估计方面的准确性,特别是在远距离或多深度区域场景中。RARE值越低,表示深度估计越准确。在远距离场景(如1.4km)和多深度分层的城市道路场景中,由于物理成像模型增强的深度约束和有效的噪声抑制,改进算法在远场深度估计方面表现出卓越的性能。实验结果表明,改进算法的相对测距误差(RARE)降低了28.2%。这意味着即使在图像质量受损或深度信息模糊的远距离区域,改进算法也能提供更为精确的距离测量,为自动驾驶和精确测绘等应用提供了更可靠的深度信息。

下表汇总了改进算法与传统方法在各项关键性能指标上的对比结果:

性能指标传统方法表现改进算法表现提升幅度关键改进策略
新颖视图合成 (PSNR)较低平均提升 5dB显著多模态数据融合、物理成像模型增强深度约束
几何重建 (Chamfer距离)较高平均降低 60%显著多视图几何约束(LoFTR)、噪声抑制(双Gamma估计)
目标恢复率 (TR)低,SBR≤0.003时失效提升 20.5%显著噪声抑制(双Gamma估计)
相对测距误差 (RARE)较高降低 28.2%显著物理成像模型增强深度约束、多视图几何约束

综上所述,改进后的三维高斯溅射算法通过融合多模态数据、引入物理成像模型和强化多视图几何与噪声抑制,在室外复杂场景下的新颖视图合成质量、几何重建精度、目标恢复率和测距精度方面均取得了突破性进展,显著优于现有传统方法。这些量化结果充分验证了改进策略的有效性和优越性。

5. 未来发展方向与挑战展望

5.1 多模态融合的泛化性提升

尽管面向室外场景的三维高斯溅射算法在改进多模态融合、物理成像模型和噪声抑制等方面取得了显著进展,但多模态融合的泛化性仍是未来研究面临的关键挑战和重要发展方向。目前的研究多集中于特定模态组合(如声纳-相机、RGB-SPAD),而在实际复杂多变且通常不可控的室外环境中,需要更自适应、更具泛化能力的多传感器融合策略,以应对更加极端的挑战。

挑战与需求:

  1. 极端天气下的跨模态数据对齐与互补:

    • 雨雪环境:在雨雪天气中,水滴和雪花会对不同传感器产生差异巨大的影响。光学相机图像可能被雨水遮挡或雪花散射,导致纹理信息丢失;激光雷达(LiDAR)的激光束会被雨滴或雪花散射和吸收,产生大量噪声和虚假点云,影响距离测量的准确性21。而热成像(Thermal Imaging)则能透过雨雪,在一定程度上提供场景的温度分布信息,但其空间分辨率通常较低,且缺乏精细的几何细节14。如何有效地融合这些在雨雪中表现各异的模态数据,实现信息的真正互补,而非简单的叠加,是未来的重点。这涉及到设计能够感知环境条件的自适应融合权重和机制,例如,在雨雪量大时,降低光学和LiDAR数据的权重,增加热成像和雷达(毫米波雷达不受雨雪影响)的权重。
    • 雾霾环境:类似地,在浓雾和重度霾环境下,可见光相机和LiDAR的性能会急剧下降,信号衰减和散射效应严重21。而热红外相机凭借其穿透烟雾的能力展现出优势14。未来的研究需要开发更智能的融合算法,能够在不同能见度条件下,动态调整各模态的贡献。
  2. 多传感器异构性:

    • 数据格式与分辨率:不同传感器输出的数据格式、分辨率、采样频率和视场(FoV)存在巨大差异。例如,LiDAR提供稀疏但精确的点云,RGB相机提供高分辨率的稠密图像,热像仪提供低分辨率的温度图。如何将这些异构数据有效对齐,并将其整合到3DGS的统一表示中,需要更复杂的配准、插值和特征提取技术。
    • 误差特性与不确定性:每种传感器都有其独特的误差来源和不确定性模型。例如,LiDAR可能受多径效应和物体反射率影响,RGB相机可能受光照变化和运动模糊影响。在融合时,需要考虑这些不确定性,并设计能够利用不确定性信息进行加权或滤波的融合策略。
  3. 自适应融合策略:

    • 动态权重分配:传统的固定权重融合方法难以应对复杂多变的室外环境。未来的研究应探索基于场景语义、环境条件或传感器可靠性评估的动态权重分配机制。例如,通过深度学习模型实时分析图像和点云特征,判断当前环境(如晴天、雨天、隧道),然后自适应地调整不同传感器模态在3DGS优化中的贡献。
    • 信息冗余与互补的平衡:在某些场景下,不同模态的信息可能存在冗余,而在另一些场景下则互为补充。理想的融合策略应能智能地识别这些情况,最大限度地利用互补信息,同时避免冗余信息带来的计算负担或冲突。

技术创新方向:

  1. 深度学习驱动的多模态融合网络:利用神经网络(如Transformer-based架构2223)来学习不同模态数据之间的复杂映射关系和内在联系,实现端到端的融合。例如,可以设计多头注意力机制,让模型自主学习如何在不同模态之间分配注意力,从而实现更鲁棒的特征融合。
  2. 概率图模型与贝叶斯融合:构建概率图模型来显式地建模不同传感器观测之间的关系及其不确定性,通过贝叶斯推断实现最优信息融合。这种方法能够更严谨地处理传感器噪声和不确定性。
  3. 场景流与动态物体建模:将多模态融合扩展到动态场景,利用场景流(scene flow)估计来处理移动物体,并将其与静态背景进行分离,从而在动态环境中实现更精确的三维重建。
  4. 知识蒸馏与迁移学习:在拥有大量标注数据的模态(如RGB)上训练模型,然后利用知识蒸馏或迁移学习将学习到的知识迁移到数据稀缺或标注困难的模态(如热成像),以增强融合模型的泛化能力。

通过不断探索这些方向,未来的多模态融合策略将能够更加自适应、更具鲁棒性,从而使三维高斯溅射算法在更广泛、更严苛的室外场景中实现高质量、高精度的三维重建。

5.2 实时性与轻量化优化

在室外场景应用中,除了重建质量和鲁棒性,实时性与轻量化是三维高斯溅射(3DGS)算法工程落地的核心需求。自动驾驶、无人机实时重建和增强现实等场景对计算资源和处理速度有着严格限制,要求算法能在有限的硬件条件下实现高效的实时处理。当前3DGS算法虽然在渲染速度上取得了显著优势,但其训练和优化过程,尤其是在大型室外场景下,仍然面临高内存消耗和计算负担的挑战 24。因此,未来研究需要重点关注参数优化的稀疏化表示和并行计算加速技术。

挑战与需求:

  1. 大规模场景下的内存消耗:一个完整的3DGS模型可能包含数百万甚至上亿个高斯点,每个高斯点都需要存储其位置、尺度、旋转、颜色(球谐系数)和不透明度等参数。在大规模室外场景中,这会导致巨大的内存占用,对部署在资源受限设备上的应用造成障碍 125。
  2. 实时优化与渲染效率:虽然3DGS的渲染速度可以达到实时甚至超实时帧率(例如,超过170 FPS在1600x900分辨率下 24),但其训练和优化过程仍然相对耗时,难以满足实时重建或动态场景更新的需求。特别是在需要频繁更新场景(如动态物体跟踪)的应用中,快速的优化迭代至关重要 2627。
  3. 计算资源限制:自动驾驶车辆和无人机等移动平台通常搭载嵌入式系统,其计算能力和功耗受到严格限制。如何在这些平台上实现高效的3DGS重建和渲染,是亟待解决的问题。

优化策略与技术创新:

  1. 参数优化的稀疏化表示:

    • 高斯剪枝与自适应密度控制:在3DGS的优化过程中,可以引入剪枝机制,移除对场景贡献较小、不透明度过低或尺寸过小的高斯点,从而减少模型复杂度。同时,结合自适应密度控制策略,在细节丰富的区域增加高斯密度,而在平坦或背景区域减少高斯密度,实现稀疏化表示 2428。例如,基于梯度的自适应稠密化机制可以根据局部梯度显著性触发高斯调整,减少冗余计算 24。
    • 量化与压缩:对高斯参数进行量化(例如,从浮点数到低位整数),可以显著减少存储空间和内存带宽需求。结合高效的压缩算法(如矢量量化、编码本技术),可以将模型大小进一步压缩,同时保持视觉质量 25。例如,注意力感知自适应编码本(AAC-GS)方法能够实现约40倍的平均压缩比,同时保持高重建质量 25。
    • 基于稀疏特征的高斯生成:探索从稀疏特征点(如SIFT、ORB特征或LiDAR点云)直接生成高斯,而非从稠密的图像像素,可以从源头减少高斯数量,降低初始模型的复杂度。
    • 混合表示:对于场景中的不同部分,采用不同的表示方法。例如,对于静态背景使用稀疏的高斯表示,而对于动态物体则采用更精细但范围较小的高斯或网格表示,从而平衡精度与效率。
  2. 并行计算加速技术:

    • GPU优化与CUDA加速:3DGS的渲染和优化过程本质上是高度并行的,非常适合在GPU上运行。通过精细的CUDA编程优化,可以最大化GPU的计算吞吐量。这包括优化内存访问模式、减少数据传输开销、利用共享内存和流多处理器(SM)等。例如,3DGS自身已经利用了CUDA加速来提高渲染效率。
    • 多核CPU与分布式计算:对于某些不适合GPU的阶段(如数据预处理、特征提取),可以利用多核CPU进行并行处理。对于超大规模场景,可以探索分布式计算架构,将场景划分为子区域,在多个计算节点上并行处理,最后再进行合并。
    • 硬件加速器与专用芯片:未来可以研究将3DGS的渲染和优化核心算法集成到FPGA或ASIC等专用硬件加速器中。例如,针对NeRF等神经渲染技术的硬件加速研究已取得进展,这些经验可以借鉴到3DGS中,实现更低功耗、更高效率的计算。
    • 实时动态场景处理:针对动态场景的实时重建,需要设计高效的时序优化策略。例如,FUGS方法通过不确定性感知变形模型和动态频谱正则化,实现了动态场景的高保真重建和实时渲染 26。

在自动驾驶与无人机实时重建中的应用:

  • 自动驾驶:通过轻量化和实时优化,3DGS可以用于生成高精度的局部地图或障碍物模型,辅助车辆进行实时感知、定位和路径规划。稀疏化表示可以减少车载计算平台的内存和计算负担,使其能在极端天气和复杂交通条件下提供稳定的三维环境感知。
  • 无人机实时重建:无人机在巡检、测绘和灾后评估等场景中,需要实时获取三维环境信息。轻量化的3DGS模型可以部署在无人机搭载的嵌入式计算单元上,实现空中实时三维重建和可视化。并行计算加速技术能确保无人机在飞行过程中,即使面对快速移动和环境变化,也能快速更新场景模型。

未来的研究将着重于将这些优化策略深度融合到3DGS框架中,并针对特定应用场景进行定制化,例如通过AI辅助设计优化轮毂结构以实现轻量化和高强度 29,这启示我们也可以利用AI设计更高效的高斯结构。同时,平衡重建质量、实时性和资源消耗之间的权衡,将是推动3DGS在室外场景广泛应用的关键。

5.3 极端环境散射模型扩展

尽管改进后的三维高斯溅射算法在低信噪比(SBR ≤ 0.003)的雾天场景中展现了卓越的性能,成功提升了目标恢复率和测距精度,但室外环境的复杂性远不止于此。未来的研究方向之一是极端环境散射模型的扩展,旨在提升单光子探测器在更低SBR(<0.003)甚至更具挑战性的散射介质(如沙尘暴、雨雪)下的三维成像能力。

现有挑战与局限性:

  1. 散射介质的多样性:目前的散射模型通常针对单一介质(如均匀的雾)进行研究。然而,室外环境中存在多种复杂的散射介质,如雾、霾、沙尘、雨、雪等,它们的散射特性(如散射系数、相函数、消光系数)各不相同,且可能随时间和空间动态变化。
  2. 物理建模的复杂性:每种散射介质的物理模型都非常复杂。例如,沙尘暴中的沙粒大小、形状、密度分布与雾中的水滴截然不同,导致光线与介质的相互作用模式发生显著变化。要准确建模这些介质对光线传输的影响,需要深入研究辐射传输理论,并结合实际测量数据。
  3. 极低SBR条件下的信号提取:在SBR低于0.003的极端条件下,信号光子几乎完全淹没在噪声和散射光子中。此时,即使是双Gamma估计等先进的噪声抑制方法也可能面临性能极限,因为信号与噪声之间的统计差异变得极其微弱。
  4. 模型泛化能力不足:现有模型通常是在特定环境条件下训练和验证的。它们可能无法直接泛化到未见过的新介质类型、密度或混合介质场景中。

改进策略与技术创新:

  1. 多介质散射模型构建与泛化:

    • 参数化散射模型:开发能够通过少量可变参数(如粒子密度、平均粒径、介质类型等)来描述不同散射介质的统一散射模型。这将允许算法在面对不同环境时,通过调整参数来适应新的散射条件,而非为每种介质从头构建模型。
    • 学习型散射模型:结合深度学习技术,利用大量不同散射环境下的模拟数据和真实数据训练神经网络,使其能够从原始单光子数据中直接学习并解耦信号、背景噪声和散射光子。这种方法有望超越传统物理模型的假设限制,捕捉更复杂的散射效应。
    • 混合散射模型:研究如何在同一场景中处理多种散射介质共存的情况(例如,地面的沙尘与高空的薄雾),构建分层或混合的散射模型。
  2. 增强型单光子探测与信号处理:

    • 量子参数模式排序(QPMS):利用量子计量学方法,如量子参数模式排序(QPMS),可以实现远超典型线性光学滤波器的信噪比,尤其在隔离时域和光谱重叠噪声方面表现突出。QPMS可以在信号光子极少(例如每脉冲仅0.0006个探测到的信号光子)的情况下,成功重建被遮挡场景的三维轮廓,即使背景噪声光子是信号光子的34倍,甚至在20ns的检测周期内背景噪声达到信号的113,000倍 30。将QPMS技术集成到单光子LiDAR系统中,有望大幅提升在极端低SBR条件下的信号提取能力。
    • 时间分辨光子探测技术:进一步优化时间分辨单光子探测技术,提高时间分辨率(ps级甚至fs级),以便更精确地分离出目标反射光子与散射光子和背景光子。更精细的时间信息可以为散射模型提供更丰富的数据,有助于解耦光子路径。
    • 结合光谱信息:除了时间信息,整合光谱信息可以为区分不同来源的光子提供额外维度。例如,利用多光谱或高光谱单光子探测器,可以根据光子的光谱特性来判断其是目标反射光、背景光还是散射光。
  3. 高斯溅射与物理模型的深度融合:

    • 物理感知的高斯表示:在3DGS的高斯参数中直接编码与散射相关的属性,例如每个高斯周围介质的散射系数。这样,在渲染过程中,可以通过物理渲染方程考虑散射效应,从而使高斯能够更准确地表示在散射介质中的外观和深度信息。
    • 迭代式散射估计与重建:采用迭代优化框架,在每次迭代中,先根据当前3DGS模型估计散射效应,然后利用估计出的散射模型对原始单光子数据进行去散射处理,再更新3DGS模型。这种自适应的迭代过程可以逐步提纯信号,改善重建质量。

通过以上策略,未来的三维高斯溅射算法有望在更恶劣的室外环境中实现稳健的三维成像,例如在极端沙尘暴中进行自动驾驶,或在深海浑浊水域进行水下探测,从而极大地扩展其应用范围。

内容由 AI 生成,仅供参考,请仔细甄别

参考文献

1Trends and Techniques in 3D Reconstruction and Rendering: A Survey with Emphasis on Gaussian Splatting.PubMed

Wenhe Chen, Zikai Li, Jingru Guo, et al.
Sensors (Basel). 2025 Jun 9;25(12):3626. doi: 10.3390/s25123626.
Three-Dimensional Gaussian Splatting (3DGS), an important advancement in the field of computer graphics and 3D vision, has emerged to greatly accelerate the rendering process in novel views' synthesis. Due to its ability to directly realize the real-time estimation of 3D shapes without neural networks, 3DGS has received a lot of attention in the fields of robotics, urban mapping, autonomous navigation, and virtual reality/augmented reality. In view of the growing popularity of 3DGS, we conduct a systematic review of the relevant literature. We begin by surveying existing work on 3D reconstruction and rendering, outlining the historical development and recent advances from both foundational and innovation-driven perspectives. Next, we summarize the most commonly used datasets and evaluation metrics in 3D reconstruction and rendering. Finally, we summarize the current challenges and suggest potential directions for future research. Through this survey, we aim to provide researchers with a treasure trove of resources in understanding and using techniques related to 3D reconstruction and rendering, in order to promote technological development and application deepening in the field of 3D vision.

2DeferredGS: Decoupled and Relightable Gaussian Splatting With Deferred Shading.PubMed

Tong Wu, Jia-Mu Sun, Yu-Kun Lai, et al.
IEEE Trans Pattern Anal Mach Intell. 2025 Aug;47(8):6307-6319. doi: 10.1109/TPAMI.2025.3560933.
Reconstructing and editing 3D objects and scenes both play crucial roles in computer graphics and computer vision. Neural radiance fields (NeRFs) can achieve realistic reconstruction and editing results but suffer from inefficiency in rendering. Gaussian splatting significantly accelerates rendering by rasterizing Gaussian ellipsoids. However, Gaussian splatting utilizes a single Spherical Harmonic (SH) function to model both texture and lighting, limiting independent editing capabilities of these components. Recently, attempts have been made to decouple texture and lighting with the Gaussian splatting representation but may fail to produce plausible geometry and decomposition results on reflective scenes. Additionally, the forward shading technique they employ introduces noticeable blending artifacts during relighting, as the geometry attributes of Gaussians are optimized under the original illumination and may not be suitable for novel lighting conditions. To address these issues, we introduce DeferredGS, a method for decoupling and relighting the Gaussian splatting representation using deferred shading. To achieve successful decoupling, we model the illumination with a learnable environment map and define additional attributes such as texture parameters and normal direction on Gaussians, where the normal is distilled from a jointly trained signed distance function. More importantly, we apply deferred shading, resulting in more realistic relighting effects compared to previous methods. Both qualitative and quantitative experiments demonstrate the superior performance of DeferredGSin novel view synthesis and relighting tasks.

3iVR-GS: Inverse Volume Rendering for Explorable Visualization via Editable 3D Gaussian Splatting.PubMed

Kaiyuan Tang, Siyuan Yao, Chaoli Wang
IEEE Trans Vis Comput Graph. 2025 Jun;31(6):3783-3795. doi: 10.1109/TVCG.2025.3567121.
In volume visualization, users can interactively explore the three-dimensional data by specifying color and opacity mappings in the transfer function (TF) or adjusting lighting parameters, facilitating meaningful interpretation of the underlying structure. However, rendering large-scale volumes demands powerful GPUs and high-speed memory access for real-time performance. While existing novel view synthesis (NVS) methods offer faster rendering speeds with lower hardware requirements, the visible parts of a reconstructed scene are fixed and constrained by preset TF settings, significantly limiting user exploration. This article introduces inverse volume rendering via Gaussian splatting (iVR-GS), an innovative NVS method that reduces the rendering cost while enabling scene editing for interactive volume exploration. Specifically, we compose multiple iVR-GS models associated with basic TFs covering disjoint visible parts to make the entire volumetric scene visible. Each basic model contains a collection of 3D editable Gaussians, where each Gaussian is a 3D spatial point that supports real-time scene rendering and editing. We demonstrate the superior reconstruction quality and composability of iVR-GS against other NVS solutions (Plenoxels, CCNeRF, and base 3DGS) on various volume datasets. The code is available at https://github.com/TouKaienn/iVR-GS.

4Collaborative 3D Scene Reconstruction in Large Outdoor Environments Using a Fleet of Mobile Ground Robots.PubMed

John Lewis, Pedro U Lima, Meysam Basiri
Sensors (Basel). 2022 Dec 29;23(1):375. doi: 10.3390/s23010375.
Teams of mobile robots can be employed in many outdoor applications, such as precision agriculture, search and rescue, and industrial inspection, allowing an efficient and robust exploration of large areas and enhancing the operators' situational awareness. In this context, this paper describes an active and decentralized framework for the collaborative 3D mapping of large outdoor areas using a team of mobile ground robots under limited communication range and bandwidth. A real-time method is proposed that allows the sharing and registration of individual local maps, obtained from 3D LiDAR measurements, to build a global representation of the environment. A conditional peer-to-peer communication strategy is used to share information over long-range and short-range distances while considering the bandwidth constraints. Results from both real-world and simulated experiments, executed in an actual solar power plant and in its digital twin representation, demonstrate the reliability and efficiency of the proposed decentralized framework for such large outdoor operations.

5Single-Pixel Near-Infrared 3D Image Reconstruction in Outdoor Conditions.PubMed

C Osorio Quero, D Durini, J Rangel-Magdaleno, et al.
Micromachines (Basel). 2022 May 20;13(5):795. doi: 10.3390/mi13050795.
In the last decade, the vision systems have improved their capabilities to capture 3D images in bad weather scenarios. Currently, there exist several techniques for image acquisition in foggy or rainy scenarios that use infrared (IR) sensors. Due to the reduced light scattering at the IR spectra it is possible to discriminate the objects in a scene compared with the images obtained in the visible spectrum. Therefore, in this work, we proposed 3D image generation in foggy conditions using the single-pixel imaging (SPI) active illumination approach in combination with the Time-of-Flight technique (ToF) at 1550 nm wavelength. For the generation of 3D images, we make use of space-filling projection with compressed sensing (CS-SRCNN) and depth information based on ToF. To evaluate the performance, the vision system included a designed test chamber to simulate different fog and background illumination environments and calculate the parameters related to image quality.

6Single-pixel imaging: An overview of different methods to be used for 3D space reconstruction in harsh environments.PubMed

Carlos A Osorio Quero, Daniel Durini, Jose Rangel-Magdaleno, et al.
Rev Sci Instrum. 2021 Nov 1;92(11):111501. doi: 10.1063/5.0050358.
Different imaging solutions have been proposed over the last few decades, aimed at three-dimensional (3D) space reconstruction and obstacle detection, either based on stereo-vision principles using active pixel sensors operating in the visible part of the spectra or based on active Near Infra-Red (NIR) illumination applying the time-of-flight principle, to mention just a few. If extremely low quantum efficiencies for NIR active illumination yielded by silicon-based detector solutions are considered together with the huge photon noise levels produced by the background illumination accompanied by Rayleigh scattering effects taking place in outdoor applications, the operating limitations of these systems under harsh weather conditions, especially if relatively low-power active illumination is used, are evident. If longer wavelengths for active illumination are applied to overcome these issues, indium gallium arsenide (InGaAs)-based photodetectors become the technology of choice, and for low-cost solutions, using a single InGaAs photodetector or an InGaAs line-sensor becomes a promising choice. In this case, the principles of Single-Pixel Imaging (SPI) and compressive sensing acquire a paramount importance. Thus, in this paper, we review and compare the different SPI developments reported. We cover a variety of SPI system architectures, modulation methods, pattern generation and reconstruction algorithms, embedded system approaches, and 2D/3D image reconstruction methods. In addition, we introduce a Near Infra-Red Single-Pixel Imaging (NIR-SPI) sensor aimed at detecting static and dynamic objects under outdoor conditions for unmanned aerial vehicle applications.

7GaussNav: Gaussian Splatting for Visual Navigation.PubMed

Xiaohan Lei, Min Wang, Wengang Zhou, et al.
IEEE Trans Pattern Anal Mach Intell. 2025 May;47(5):4108-4121. doi: 10.1109/TPAMI.2025.3538496. Epub 2025 Apr 8.
In embodied vision, Instance ImageGoal Navigation (IIN) requires an agent to locate a specific object depicted in a goal image within an unexplored environment. The primary challenge of IIN arises from the need to recognize the target object across varying viewpoints while ignoring potential distractors. Existing map-based navigation methods typically use Bird's Eye View (BEV) maps, which lack detailed texture representation of a scene. Consequently, while BEV maps are effective for semantic-level visual navigation, they are struggling for instance-level tasks. To this end, we propose a new framework for IIN, Gaussian Splatting for Visual Navigation (GaussNav), which constructs a novel map representation based on 3D Gaussian Splatting (3DGS). The GaussNav framework enables the agent to memorize both the geometry and semantic information of the scene, as well as retain the textural features of objects. By matching renderings of similar objects with the target, the agent can accurately identify, ground, and navigate to the specified object. Our GaussNav framework demonstrates a significant performance improvement, with Success weighted by Path Length (SPL) increasing from 0.347 to 0.578 on the challenging Habitat-Matterport 3D (HM3D) dataset.

83D Gaussian Splatting as a New Era: A Survey.PubMed

Ben Fei, Jingyi Xu, Rui Zhang, et al.
IEEE Trans Vis Comput Graph. 2025 Aug;31(8):4429-4449. doi: 10.1109/TVCG.2024.3397828.
3D Gaussian Splatting (3D-GS) has emerged as a significant advancement in the field of computer graphics and 3D vision, offering explicit scene representation and novel view synthesis without the reliance on neural networks. This technique has found diverse applications in areas such as robotics, urban mapping, autonomous navigation, and virtual reality/augmented reality, just name a few. Given the growing popularity and expanding research in 3D-GS, this paper presents a comprehensive survey of relevant papers from the past year. We organize the survey into taxonomies based on characteristics and applications, providing an introduction to the theoretical underpinnings of 3D-GS. The survey aims to introduce the theoretical foundations of 3D Gaussian Splatting and provide a reference for new researchers while inspiring future research directions.

9Three-Dimensional Environment Mapping with a Rotary-Driven Lidar in Real Time.PubMed

Baixin Tong, Fangdi Jiang, Bo Lu, et al.
Sensors (Basel). 2025 Aug 7;25(15):4870. doi: 10.3390/s25154870.
Three-dimensional environment reconstruction refers to the creation of mathematical models of three-dimensional objects suitable for computer representation and processing. This paper proposes a novel 3D environment reconstruction approach that addresses the field-of-view limitations commonly faced by LiDAR-based systems. A rotary-driven LiDAR mechanism is designed to enable uniform and seamless full-field-of-view scanning, thereby overcoming blind spots in traditional setups. To complement the hardware, a multi-sensor fusion framework-LV-SLAM (LiDAR-Visual Simultaneous Localization and Mapping)-is introduced. The framework consists of two key modules: multi-threaded feature registration and a two-phase loop closure detection mechanism, both designed to enhance the system's accuracy and robustness. Extensive experiments on the KITTI benchmark demonstrate that LV-SLAM outperforms state-of-the-art methods including LOAM, LeGO-LOAM, and FAST-LIO2. Our method reduces the average absolute trajectory error (ATE) from 6.90 m (LOAM) to 2.48 m, and achieves lower relative pose error (RPE), indicating improved global consistency and reduced drift. We further validate the system in real-world indoor and outdoor environments. Compared with fixed-angle scans, the rotary LiDAR mechanism produces more complete reconstructions with fewer occlusions. Geometric accuracy evaluation shows that the root mean square error between reconstructed and actual building dimensions remains below 5 cm. The proposed system offers a robust and accurate solution for high-fidelity 3D reconstruction, particularly suitable for GNSS-denied and structurally complex environments.

10Real-time 3D reconstruction from single-photon lidar data using plug-and-play point cloud denoisers.PubMed

Julián Tachella, Yoann Altmann, Nicolas Mellado, et al.
Nat Commun. 2019 Nov 1;10(1):4984. doi: 10.1038/s41467-019-12943-7.
Single-photon lidar has emerged as a prime candidate technology for depth imaging through challenging environments. Until now, a major limitation has been the significant amount of time required for the analysis of the recorded data. Here we show a new computational framework for real-time three-dimensional (3D) scene reconstruction from single-photon data. By combining statistical models with highly scalable computational tools from the computer graphics community, we demonstrate 3D reconstruction of complex outdoor scenes with processing times of the order of 20 ms, where the lidar data was acquired in broad daylight from distances up to 320 metres. The proposed method can handle an unknown number of surfaces in each pixel, allowing for target detection and imaging through cluttered scenes. This enables robust, real-time target reconstruction of complex moving scenes, paving the way for single-photon lidar at video rates for practical 3D imaging applications.

11Multi-scale geometric transformer for sparse-view X-ray 3D foot reconstruction.PubMed

Wei Wang, Li An, Gengyin Han
J Xray Sci Technol. 2025 Jul;33(4):776-787. doi: 10.1177/08953996251319194. Epub 2025 Apr 25.
BACKGROUND: Sparse-View X-ray 3D Foot Reconstruction aims to reconstruct the three-dimensional structure of the foot from sparse-view X-ray images, a challenging task due to data sparsity and limited viewpoints. OBJECTIVE: This paper presents a novel method using a multi-scale geometric Transformer to enhance reconstruction accuracy and detail representation. METHODS: Geometric position encoding technology and a window mechanism are introduced to divide X-ray images into local areas, finely capturing local features. A multi-scale Transformer module based on Neural Radiance Fields (NeRF) enhances the model's ability to express and capture details in complex structures. An adaptive weight learning strategy further optimizes the Transformer's feature extraction and long-range dependency modelling. RESULTS: Experimental results demonstrate that the proposed method significantly improves the reconstruction accuracy and detail preservation of the foot structure under sparse-view X-ray conditions. The multi-scale geometric Transformer effectively captures local and global features, leading to more accurate and detailed 3D reconstructions. CONCLUSIONS: The proposed method advances medical image reconstruction, significantly improving the accuracy and detail preservation of 3D foot reconstructions from sparse-view X-ray images.

12SERV-CT: A disparity dataset from cone-beam CT for validation of endoscopic 3D reconstruction.PubMed

P J Eddie Edwards, Dimitris Psychogyios, Stefanie Speidel, et al.
Med Image Anal. 2022 Feb;76:102302. doi: 10.1016/j.media.2021.102302. Epub 2021 Nov 6.
In computer vision, reference datasets from simulation and real outdoor scenes have been highly successful in promoting algorithmic development in stereo reconstruction. Endoscopic stereo reconstruction for surgical scenes gives rise to specific problems, including the lack of clear corner features, highly specular surface properties and the presence of blood and smoke. These issues present difficulties for both stereo reconstruction itself and also for standardised dataset production. Previous datasets have been produced using computed tomography (CT) or structured light reconstruction on phantom or ex vivo models. We present a stereo-endoscopic reconstruction validation dataset based on cone-beam CT (SERV-CT). Two ex vivo small porcine full torso cadavers were placed within the view of the endoscope with both the endoscope and target anatomy visible in the CT scan. Subsequent orientation of the endoscope was manually aligned to match the stereoscopic view and benchmark disparities, depths and occlusions are calculated. The requirement of a CT scan limited the number of stereo pairs to 8 from each ex vivo sample. For the second sample an RGB surface was acquired to aid alignment of smooth, featureless surfaces. Repeated manual alignments showed an RMS disparity accuracy of around 2 pixels and a depth accuracy of about 2 mm. A simplified reference dataset is provided consisting of endoscope image pairs with corresponding calibration, disparities, depths and occlusions covering the majority of the endoscopic image and a range of tissue types, including smooth specular surfaces, as well as significant variation of depth. We assessed the performance of various stereo algorithms from online available repositories. There is a significant variation between algorithms, highlighting some of the challenges of surgical endoscopic images. The SERV-CT dataset provides an easy to use stereoscopic validation for surgical applications with smooth reference disparities and depths covering the majority of the endoscopic image. This complements existing resources well and we hope will aid the development of surgical endoscopic anatomical reconstruction algorithms.

13Three-dimensional angular scattering simulations inform analysis of scattering from single cells.PubMed

Kaitlin J Dunn, Andrew J Berger
J Biomed Opt. 2023 Aug;28(8):086501. doi: 10.1117/1.JBO.28.8.086501. Epub 2023 Aug 9.
SIGNIFICANCE: Organelle sizes, which are indicative of cellular status, have implications for drug development and immunology research. At the single cell level, such information could be used to study the heterogeneity of cell response to drugs or pathogens. AIM: Angularly resolved elastic light scattering is known to be sensitive to changes in organelle size distribution. We developed a Mie theory-based simulation of angular scattering from single cells to quantify the effects of noise on scattering and size estimates. APPROACH: We simulated randomly sampled organelle sizes (drawn from a log normal distribution), interference between different organelles' scattering, and detector noise. We quantified each noise source's effect upon the estimated mean and standard deviation of organelle size distributions. RESULTS: The results demonstrate that signal-to-noise ratio in the angular scattering increased with the number of scatterers, cell area, and exposure time and decreased with the size distribution width. The error in estimating the mean of the size distributions remained below 5% for nearly all experimental parameters tested, but the widest size distribution tested (standard deviation of 600 nm) reached 20%. CONCLUSIONS: The simulator revealed that sparse sampling of a broad size distribution can dominate the mismatch between actual and predicted size parameters. Alternative estimation strategies could reduce the discrepancy.

14Thermal3D-GS: Physics-induced 3D Gaussians for Thermal Infrared Novel-view Synthesis with a Large-Scale Dataset.PubMed

Qian Chen, Shihao Shu, Heng Sun, et al.
IEEE Trans Pattern Anal Mach Intell. 2026 Feb 11;PP. doi: 10.1109/TPAMI.2026.3663966.
Thermal infrared imaging has attracted widespread attention in many fields due to the advantages of all-weather imaging and strong penetration. However, existing methods for thermal infrared novel-view synthesis often produce results with coarse details and floating artifacts, primarily caused by physical factors such as atmospheric transmission effects and thermal conduction. These challenges hinder accurate reconstruction of intricate structures and temperature distributions in thermal scenes, limiting the practical utility of previous approaches. To address these limitations, this paper introduces a physics-induced 3D Gaussian splatting method named Thermal3D-GS, the first novel-view synthesis method that relies exclusively on thermal infrared image. Thermal3D-GS begins by modeling atmospheric transmission effects and thermal conduction in three-dimensional media using neural networks. Additionally, considering the sparse features of infrared images, sparse feature priors are designed to improve the reconstruction accuracy of thermal infrared images. Furthermore, to validate the effectiveness of our method, the first large-scale benchmark dataset named Thermal Infrared Novel-view Synthesis Dataset (TI-NSD) is created. This dataset comprises 50 authentic thermal infrared video scenes, covering indoor, outdoor, traffic and UAV(Unmanned Aerial Vehicle) scenarios, with a total of 15,213 frames of thermal infrared image data. In addition, an expanded validation thermal infrared dataset, which includes three high-resolution scenes and five special scenes under varying atmospheric conditions and complex propagation media is constructed to assess generalization performance of the proposed method. Based on this dataset, this paper experimentally verifies the effectiveness of Thermal3D-GS. The results indicate that our method outperforms the baseline method with a 3.19 dB improvement in PSNR and significantly addresses the issues of floaters and indistinct edge features present in the baseline method. The dataset and our code are both publicly available in https://github.com/mzzcdf/Thermal3DGS.

15The darker-is-deeper heuristic for the perception of 3D shape from shading: Is it perceptually or ecologically valid?PubMed

James T Todd, Eric J L Egan, Christopher S Kallie
J Vis. 2015;15(15):2. doi: 10.1167/15.15.2.
The darker-is-deeper heuristic was originally proposed by Langer and Zucker (1994) for approximating 3D shape from shading under conditions of diffuse illumination that typically occur for outdoor scenes on a cloudy day, and it is based on the assumption that vignetting is the primary source of luminance variation under those conditions. It was later rejected as a model of human perception by Langer and Bülthoff (2000), because points in concavities that appear to be the deepest are most often located on local luminance maxima. Despite that result, this heuristic has continued to be described in the literature as a viable model of human perception (e.g., Chen & Tyler, 2015; Tyler, 1998), based entirely on the appearance of image intensity gratings, which have little or no connection to real 3D surfaces or patterns of illumination. In this article we will present a large number of examples to show what actually happens when surfaces are viewed under directional and diffuse illuminations. The results will highlight a number of well-known phenomena in addition to vignetting that can influence the pattern of shading on a surface under diffuse illumination, and they will also demonstrate that the darker-is-deeper heuristic is generally invalid for all types of illumination, except in unusual circumstances.

16Depth-consistent 3D Gaussian Splatting via physical defocus modeling and multi-view geometric supervision.PubMed

Yu Deng, Baozhu Zhao, Junyan Su, et al.
Neural Netw. 2026 Mar;195:108320. doi: 10.1016/j.neunet.2025.108320. Epub 2025 Nov 13.
Three-dimensional reconstruction in scenes with extreme depth variations remains challenging due to inconsistent supervisory signals between near-field and far-field regions. Existing methods fail to simultaneously address inaccurate depth estimation in distant areas and structural degradation in close-range regions. This paper proposes a novel computational framework that integrates depth-of-field supervision and multi-view consistency supervision to advance 3D Gaussian Splatting. Our approach comprises two core components: (1) Depth-of-field Supervision employs a scale-recovered monocular depth estimator (e.g., Metric3D) to generate depth priors, leverages defocus convolution to synthesize physically accurate defocused images, and enforces geometric consistency through a novel depth-of-field loss, thereby enhancing depth fidelity in both far-field and near-field regions; (2) Multi-View Consistency Supervision employing LoFTR-based semi-dense feature matching to minimize cross-view geometric errors and enforce depth consistency via least squares optimization of reliable matched points. By unifying defocus physics with multi-view geometric constraints, our method achieves superior depth fidelity, demonstrating a 0.8 dB PSNR improvement over the state-of-the-art method on the Waymo Open Dataset. This framework bridges physical imaging principles and learning-based depth regularization, offering a scalable solution for complex depth stratification in urban environments.

17Detection of Simulated Fukushima Daichii Fuel Debris Using a Remotely Operated Vehicle at the Naraha Test Facility.PubMed

Matthew Nancekievill, Jose Espinosa, Simon Watson, et al.
Sensors (Basel). 2019 Oct 22;19(20):4602. doi: 10.3390/s19204602.
The use of robotics in harsh environments, such as nuclear decommissioning, has increased in recent years. Environments such as the Fukushima Daiichi accident site from 2011 and the Sellafield legacy ponds highlight the need for robotic systems capable of deployment in hazardous environments unsafe for human workers. To characterise these environments, it is important to develop robust and accurate localization systems that can be combined with mapping techniques to create 3D reconstructions of the unknown environment. This paper describes the development and experimental verification of a localization system for an underwater robot, which enabled the collection of sonar data to create 3D images of submerged simulated fuel debris. The system was demonstrated at the Naraha test facility, Fukushima prefecture, Japan. Using a camera with a bird's-eye view of the simulated primary containment vessel, the 3D position and attitude of the robot was obtained using coloured LED markers (active markers) on the robot, landmarks on the test-rig (passive markers), and a depth sensor on the robot. The successful reconstruction of a 3D image has been created through use of a robot operating system (ROS) node in real-time.

18Three-dimensional single-photon imaging through realistic fog in an outdoor environment during the day.PubMed

Yinbo Zhang, Sining Li, Jianfeng Sun, et al.
Opt Express. 2022 Sep 12;30(19):34497-34509. doi: 10.1364/OE.464297.
Due to the strong scattering of fog and the strong background noise, the signal-to-background ratio (SBR) is extremely low, which severely limits the 3D imaging capability of single-photon detector array through fog. Here, we propose an outdoor three-dimensional imaging algorithm through fog, which can separate signal photons from non-signal photons (scattering and noise photons) with SBR as low as 0.003. This is achieved by using the observation model based on multinomial distribution to compensate for the pile-up, and using dual-Gamma estimation to eliminate non-signal photons. We show that the proposed algorithm enables accurate 3D imaging of 1.4 km in the visibility of 1.7 km. Compared with the traditional algorithms, the target recovery (TR) of the reconstructed image is improved by 20.5%, and the relative average ranging error (RARE) is reduced by 28.2%. It has been successfully demonstrated for targets at different distances and imaging times. This research successfully expands the fog scattering estimation model from indoor to outdoor environment, and improves the weather adaptability of the single-photon detector array.

19Comparative Assessment of Neural Radiance Fields and 3D Gaussian Splatting for Point Cloud Generation from UAV Imagery.PubMed

Muhammed Enes Atik
Sensors (Basel). 2025 May 9;25(10):2995. doi: 10.3390/s25102995.
Point clouds continue to be the main data source in 3D modeling studies with unmanned aerial vehicle (UAV) images. Structure-from-Motion (SfM) and MultiView Stereo (MVS) have high time costs for point cloud generation, especially in large data sets. For this reason, state-of-the-art methods such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have emerged as powerful alternatives for point cloud generation. This paper explores the performance of NeRF and 3DGS methods in generating point clouds from UAV images. For this purpose, the Nerfacto, Instant-NGP, and Splatfacto methods developed in the Nerfstudio framework were used. The obtained point clouds were evaluated by taking the point cloud produced with the photogrammetric method as reference. In this study, the effects of image size and iteration number on the performance of the algorithms were investigated in two different study areas. According to the results, Splatfacto demonstrates promising capabilities in addressing challenges related to scene complexity, rendering efficiency, and accuracy in UAV imagery.

20Research on the Depth Image Reconstruction Algorithm Using the Two-Dimensional Kaniadakis Entropy Threshold.PubMed

Xianhui Yang, Jianfeng Sun, Le Ma, et al.
Sensors (Basel). 2024 Sep 13;24(18):5950. doi: 10.3390/s24185950.
The photon-counting light laser detection and ranging (LiDAR), especially the Geiger mode avalanche photon diode (Gm-APD) LiDAR, can obtain three-dimensional images of the scene, with the characteristics of single-photon sensitivity, but the background noise limits the imaging quality of the laser radar. In order to solve this problem, a depth image estimation method based on a two-dimensional (2D) Kaniadakis entropy thresholding method is proposed which transforms a weak signal extraction problem into a denoising problem for point cloud data. The characteristics of signal peak aggregation in the data and the spatio-temporal correlation features between target image elements in the point cloud-intensity data are exploited. Through adequate simulations and outdoor target-imaging experiments under different signal-to-background ratios (SBRs), the effectiveness of the method under low signal-to-background ratio conditions is demonstrated. When the SBR is 0.025, the proposed method reaches a target recovery rate of 91.7%, which is better than the existing typical methods, such as the Peak-picking method, Cross-Correlation method, and the sparse Poisson intensity reconstruction algorithm (SPIRAL), which achieve a target recovery rate of 15.7%, 7.0%, and 18.4%, respectively. Additionally, comparing with the SPIRAL, the reconstruction recovery ratio is improved by 73.3%. The proposed method greatly improves the integrity of the target under high-background-noise environments and finally provides a basis for feature extraction and target recognition.

213D object detection through fog and occlusion: passive integral imaging vs active (LiDAR) sensing.PubMed

Kashif Usmani, Timothy O'Connor, Pranav Wani, et al.
Opt Express. 2023 Jan 2;31(1):479-491. doi: 10.1364/OE.478125.
In this paper, we address the problem of object recognition in degraded environments including fog and partial occlusion. Both long wave infrared (LWIR) imaging systems and LiDAR (time of flight) imaging systems using Azure Kinect, which combine conventional visible and lidar sensing information, have been previously demonstrated for object recognition in ideal conditions. However, the object detection performance of Azure Kinect depth imaging systems may decrease significantly in adverse weather conditions such as fog, rain, and snow. The concentration of fog degrades the depth images of Azure Kinect camera, and the overall visibility of RGBD images (fused RGB and depth image), which can make object recognition tasks challenging. LWIR imaging may avoid these issues of lidar-based imaging systems. However, due to poor spatial resolution of LWIR cameras, thermal imaging provides limited textural information within a scene and hence may fail to provide adequate discriminatory information to identify between objects of similar texture, shape and size. To improve the object detection task in fog and occlusion, we use three-dimensional (3D) integral imaging (InIm) system with a visible range camera. 3D InIm provides depth information, mitigates the occlusion and fog in front of the object, and improves the object recognition capabilities. For object recognition, the YOLOv3 neural network is used for each of the tested imaging systems. Since the concentration of fog affects the images from different sensors (visible, LWIR, and Azure Kinect depth cameras) in different ways, we compared the performance of the network on these images in terms of average precision and average miss rate. For the experiments we conducted, the results indicate that in degraded environment 3D InIm using visible range cameras can provide better image reconstruction as compared to the LWIR camera and Azure Kinect RGBD camera, and therefore it may improve the detection accuracy of the network. To the best of our knowledge, this is the first report comparing the performance of object detection between passive integral imaging system vs active (LiDAR) sensing in degraded environments such as fog and partial occlusion.

22JM3D & JM3D-LLM: Elevating 3D Representation With Joint Multi-Modal Cues.PubMed

Jiayi Ji, Haowei Wang, Changli Wu, et al.
IEEE Trans Pattern Anal Mach Intell. 2025 Apr;47(4):2475-2492. doi: 10.1109/TPAMI.2024.3523675. Epub 2025 Mar 6.
The rising importance of 3D representation learning, pivotal in computer vision, autonomous driving, and robotics, is evident. However, a prevailing trend, which straightforwardly resorted to transferring 2D alignment strategies to the 3D domain, encounters three distinct challenges: (1) Information Degradation: This arises from the alignment of 3D data with mere single-view 2D images and generic texts, neglecting the need for multi-view images and detailed subcategory texts. (2) Insufficient Synergy: These strategies align 3D representations to image and text features individually, hampering the overall optimization for 3D models. (3) Underutilization: The fine-grained information inherent in the learned representations is often not fully exploited, indicating a potential loss in detail. To address these issues, we introduce JM3D, a comprehensive approach integrating point cloud, text, and image. Key contributions include the Structured Multimodal Organizer (SMO), enriching vision-language representation with multiple views and hierarchical text, and the Joint Multi-modal Alignment (JMA), combining language understanding with visual representation. Our advanced model, JM3D-LLM, marries 3D representation with large language models via efficient fine-tuning. Evaluations on ModelNet40 and ScanObjectNN establish JM3D's superiority. The superior performance of JM3D-LLM further underscores the effectiveness of our representation transfer approach.

23MHAFR-DDI: a multimodal hierarchical attention fusion and relation-aware architecture for drug-drug interaction event prediction.PubMed

Mengli Li, Chao Cao, Quan Zou, et al.
Brief Bioinform. 2026 Jan 7;27(1). doi: 10.1093/bib/bbag077.
Drug-drug interactions (DDIs) can lead to severe adverse reactions, and accurate prediction of DDI events is crucial for ensuring the safety of combination therapies and supporting drug development. Although deep learning-based approaches have achieved promising progress, existing models remain limited in modeling local-global dependencies, integrating multimodal information, and capturing cross-level molecular relationships. To address these challenges, we propose Multi-modal Hierarchical Attention Fusion and Relation-aware Architecture for DDI Event Prediction (MHAFR-DDI), a multimodal hierarchical attention fusion and relation-aware framework that enables unified modeling from intra-molecular representation to inter-molecular interaction. MHAFR-DDI adopts a two-stage pretraining-finetuning paradigm. In the pretraining stage, the model learns complementary representations from molecular sequences, 2D topological structures, and 3D spatial conformations, with modality-specific encoding mechanisms designed to capture both local structural characteristics and global semantic dependencies. Localized chemical primitives within each modality are first stabilized and then integrated into higher-level representations to ensure intra-modality stability and representational completeness. Subsequently, by introducing attention-guided data augmentation and multi-level contrastive learning, the model establishes alignment constraints across different modalities and their augmented views, thereby achieving cross-modal semantic consistency and effectively alleviating data sparsity. During the finetuning stage, the pretrained molecular representations are hierarchically fused and propagated over the drug-drug interaction graph, enabling interaction-aware information sharing among drugs and improving prediction reliability for rare drugs and long-tail interaction types. Experiments on benchmarks with 65 and 86 DDI types show that MHAFR-DDI outperforms state-of-the-art methods under the standard split, achieving macro-F1 gains of 9.5% and 6.7%, while remaining robust in weakly supervised long-tail and cold-start settings.

24Enhanced 3D Gaussian Splatting for Real-Scene Reconstruction via Depth Priors, Adaptive Densification, and Denoising.PubMed

Haixing Shang, Mengyu Chen, Kenan Feng, et al.
Sensors (Basel). 2025 Nov 16;25(22):6999. doi: 10.3390/s25226999.
The application prospects of photorealistic 3D reconstruction are broad in smart cities, cultural heritage preservation, and related domains. However, existing methods face persistent challenges in balancing reconstruction accuracy, computational efficiency, and robustness, particularly in complex scenes characterized by reflective surfaces, vegetation, sparse viewpoints, or large-scale structures. In this study, an enhanced 3D Gaussian Splatting (3DGS) framework that integrates three key innovations is proposed: (i) a depth-aware regularization module that leverages metric depth priors from the pre-trained Depth-Anything V2 model, enabling geometrically informed optimization through a dynamically weighted hybrid loss; (ii) a gradient-driven adaptive densification mechanism that triggers Gaussian adjustments based on local gradient saliency, reducing redundant computation; and (iii) a neighborhood density-based floating artifact detection method that filters outliers using spatial distribution and opacity thresholds. Extensive evaluations are conducted across four diverse datasets-ranging from architectures, urban scenes, natural landscapes with water bodies, and long-range linear infrastructures. Our method achieves state-of-the-art performance in both reconstruction quality and efficiency, attaining a PSNR of 34.15 dB and SSIM of 0.9382 on medium-sized scenes, with real-time rendering speeds exceeding 170 FPS at a resolution of 1600 × 900. It demonstrates superior generalization on challenging materials such as water and foliage, while exhibiting reduced overfitting compared to baseline approaches. Ablation studies confirm the critical contributions of depth regularization and gradient-sensitive adaptation, with the latter improving training efficiency by 38% over depth supervision alone. Furthermore, we analyze the impact of input resolution and depth model selection, revealing non-trivial trade-offs between quantitative metrics and visual fidelity. While aggressive downsampling inflates PSNR and SSIM, it leads to loss of high-frequency detail; we identify 1/4-1/2 resolution scaling as an optimal balance for practical deployment. Among depth models, Vitb achieves the best reconstruction stability. Despite these advances, memory consumption remains a challenge in large-scale scenarios. Future work will focus on lightweight model design, efficient point cloud preprocessing, and dynamic memory management to enhance scalability for industrial applications.

25AAC-GS: Attention-aware adaptive codebook for Gaussian splatting compression.PubMed

Fang Wan, Jianhang Zhang, Tianyu Li, et al.
Neural Netw. 2026 Feb;194:108134. doi: 10.1016/j.neunet.2025.108134. Epub 2025 Sep 20.
Neural Radiance Fields (NeRF) have demonstrated remarkable performance in the field of novel view synthesis (NVS). However, their high computational cost limits practical applicability. The 3D Gaussian Splatting (3DGS) method offers a significant improvement in rendering efficiency, enabling real-time rendering through its explicit representations. Nevertheless, its substantial storage requirements pose challenges for complex scenes and resource-constrained devices. Existing methods aim to achieve storage compression through redundant point pruning, spherical harmonics adjustment, and vector quantization. However, point pruning methods often compromise geometric details in complex structures, while vector quantization approaches fail to capture feature relationships effectively, resulting in texture degradation and geometric boundary blurring. Although anchor point representations partially address storage concerns, their sparse representation limits compression efficiency. These limitations become particularly evident in scenes with intricate textures and complex lighting conditions. To ensure optimal compression ratios while maintaining high fidelity in Gaussian scenarios, this paper proposes an Attention-Aware Adaptive Codebook Gaussian Splatting (AAC-GS) method for efficient storage compression. The approach dynamically adjusts the size of the codebook to optimize storage efficiency and incorporates an attention mechanism to capture feature contextual relationships, thereby enhancing reconstruction quality. Additionally, a Generative Adversarial Network (GAN) is employed to mitigate quantization losses, achieving a balance between compression rate and visual fidelity. Experimental results demonstrate that AAC-GS achieves an average compression ratio of approximately 40× while maintaining high reconstruction quality, showcasing its potential for multi-scene applications.

26Frequency-Aware Uncertainty Gaussian Splatting for Dynamic Scene Reconstruction.PubMed

Mingwen Shao, Yuanjian Qiao, Kai Zhang, et al.
IEEE Trans Vis Comput Graph. 2025 May;31(5):3558-3568. doi: 10.1109/TVCG.2025.3549143. Epub 2025 Apr 25.
3D Gaussian splatting has recently achieved remarkable progress in dynamic scene reconstruction. However, there remain two practical challenges: (1) Existing methods typically employ a strict point-wise deformation structure to model dynamic attributes, while neglecting the uncertain motion correlation in local space, leading to inferior adaptability to complex scenes. (2) The inherent low-frequency bias properties of Gaussians often lead to blurring artifacts due to the insufficient high-frequency learning of variable motions. To address these challenges, we propose a novel Frequency-aware Uncertainty Gaussian Splatting, termed FUGS, for adaptively reconstructing dynamic scenes in the Fourier space. Specifically, we design an Uncertainty-aware Deformation Model (UDM) that explicitly models motion attributes using learnable uncertainty relations with neighboring Gaussian points. Such a paradigm is capable of facilitating temporal and spatial motion correlation learning, thereby enabling flexible Gaussian deformations. Subsequently, a Dynamic Spectrum Regularization (DSR) is developed to perform coarse-to-fine Gaussian densification through low-to-high frequency filtering. By weighting the gradient with frequency distance, the Gaussian attribute is adaptively adjusted according to the scene complexity. Benefiting from the flexible optimization, our method achieves high-fidelity reconstruction of complex scenes while enjoying real-time rendering. Extensive experiments on synthetic and real-world datasets show that our FUGS exhibits significant superiority over state-of-the-art methods. The code will be available at https://github.com/KevinJoee/GS.

27Foundation Model-Guided Gaussian Splatting for 4D Reconstruction of Deformable Tissues.PubMed

Yifan Liu, Chenxin Li, Hengyu Liu, et al.
IEEE Trans Med Imaging. 2025 Jun;44(6):2672-2682. doi: 10.1109/TMI.2025.3545183.
Reconstructing deformable anatomical structures from endoscopic videos is a pivotal and promising research topic that can enable advanced surgical applications and improve patient outcomes. While existing surgical scene reconstruction methods have made notable progress, they often suffer from slow rendering speeds due to using neural radiance fields, limiting their practical viability in real-world applications. To overcome this bottleneck, we propose EndoGaussian, a framework that integrates the strengths of 3D Gaussian Splatting representations, allowing for high-fidelity tissue reconstruction, efficient training, and real-time rendering. Specifically, we dedicate a Foundation Model-driven Initialization (FMI) module, which distills 3D cues from multiple vision foundation models (VFMs) to swiftly construct the preliminary scene structure for Gaussian initialization. Then, a Spatio-temporal Gaussian Tracking (SGT) is designed, efficiently modeling scene dynamics using the multi-scale HexPlane with spatio-temporal priors. Furthermore, to improve the dynamics modeling ability for scenes with large deformation, EndoGaussian integrates Motion-aware Frame Synthesis (MFS) to adaptively synthesize new frames as extra training constraints. Experimental results on public datasets demonstrate EndoGaussian's efficacy against prior state-of-the-art methods, including superior rendering speed (168 FPS, real-time), enhanced rendering quality (38.555 PSNR), and reduced training overhead (within 2 min/scene). These results underscore EndoGaussian's potential to significantly advance intraoperative surgery applications, paving the way for more accurate and efficient real-time surgical guidance and decision-making in clinical scenarios. Code is available at: https://github.com/CUHK-AIM-Group/EndoGaussian.

28SeG-Gaussian:Segmentation-Guided 3D Gaussian Optimization for Novel View Synthesis.PubMed

Ling-Xiao Zhang, Chenbo Jiang, Yu-Kun Lai, et al.
IEEE Trans Vis Comput Graph. 2025 Dec;31(12):10815-10827. doi: 10.1109/TVCG.2025.3615421.
Radiance field based methods have recently revolutionized novel view synthesis of scenes captured with multi-view photos. A significant recent advance is 3D Gaussian Splatting (3DGS), which utilizes a set of 3D Gaussians to represent a radiance field, yielding high-fidelity results in real-time rendering. However, we have observed that 3DGS struggles to capture the necessary details in sparsely observed regions, where there is not enough gradient for effective split and clone operations. In this paper, we present a novel solution to address this limitation. Our key idea is to leverage segmentation information to identify poorly optimized regions within the 3D Gaussian representation. By applying split or clone operations on the corresponding 3D Gaussians in these regions, we aim to refine the spatial distribution of Gaussians and enhance the overall quality of high-fidelity 3D scene reconstruction. To further optimize the reconstruction process, we introduce two spatial regularization terms: repulsion loss and smoothness loss. These terms effectively minimize overlap and redundancy among Gaussians, reducing outliers in the synthesized geometry. By incorporating these regularization techniques, our approach achieves state-of-the-art performance in real-time novel view synthesis and significantly improves visibility in less observed regions, leading to a more compact and accurate 3D scene representation.

29AI-assisted design of lightweight and strong 3D-printed wheels for electric vehicles.PubMed

Timileyin Opeyemi Akande, Oluwaseyi O Alabi, Ali Rizwan, et al.
PLoS One. 2024 Dec 2;19(12):e0308004. doi: 10.1371/journal.pone.0308004. eCollection 2024.
The automotive industry is undergoing a transformative shift towards electric vehicles (EVs), driven by environmental concerns and technological advancements. One critical aspect of EV design is the development of lightweight yet robust components, including 3D vehicle wheels. This research explores the implementation of generative models in Computer-Aided Design (CAD) systems to optimize the design of 3D vehicle wheels for electric vehicles. Through the use of generative design and additive manufacturing, we aim to create vehicle wheels that are energy-efficient, aesthetically pleasing, and structurally sound. Electric vehicles are gaining popularity due to their environmental benefits and reduced operating costs, making lightweight and strong wheels an important design goal. This research proposes a novel approach for designing lightweight and strong 3D vehicle wheels for EVs using generative models. The proposed approach involves the following steps: collect and prepare data, choose a generative model architecture, train the generative model, and generate new wheel designs. The approach methods show potential to revolutionize the design and manufacturing of lightweight and strong 3D-printed wheels for electric vehicles. In conclusion, generative models can be used to design and optimize wheel designs, making it possible to create safer, more efficient, and more cost-effective wheels.

30Noise-tolerant single photon sensitive three-dimensional imager.PubMed

Patrick Rehain, Yong Meng Sua, Shenyu Zhu, et al.
Nat Commun. 2020 Feb 17;11(1):921. doi: 10.1038/s41467-020-14591-8.
Active imagers capable of reconstructing 3-dimensional (3D) scenes in the presence of strong background noise are highly desirable for many sensing and imaging applications. A key to this capability is the time-resolving photon detection that distinguishes true signal photons from the noise. To this end, quantum parametric mode sorting (QPMS) can achieve signal to noise exceeding by far what is possible with typical linear optics filters, with outstanding performance in isolating temporally and spectrally overlapping noise. Here, we report a QPMS-based 3D imager with exceptional detection sensitivity and noise tolerance. With only 0.0006 detected signal photons per pulse, we reliably reconstruct the 3D profile of an obscured scene, despite 34-fold spectral-temporally overlapping noise photons, within the 6 ps detection window (amounting to 113,000 times noise per 20 ns detection period). Our results highlight a viable approach to suppress background noise and measurement errors of single photon imager operation in high-noise environments.