2021-2023神经辐射场（NeRF）118篇顶会论文合集（已分类整理）

有没有发现最近神经辐射场相关话题又有点爆了？

其实NeRF一直都挺火的，相关论文在Google Scholar上引用量已经超过6000次，每年顶级视觉会议如CVPR、ICCV、ECCV等都有大量关于神经辐射场的高质量工作被接受。许多科研机构和公司比如微软、Facebook、谷歌等都在研究神经辐射场的不同应用，有的已经将其应用于产品中，如捕捉人体动作等。

今年有关NeRF的最新研究也有不少了，上次和你们简单分享了17篇，传送门点这里。

这次我爆肝整理了2021-2023近三年神经辐射场相关的顶会论文118篇，想发paper的同学务必收藏，抓紧时间学起来了。

论文资料已经帮大家做了分类，包含综述、姿态估计、推理、训练、组合性、场景标注、对象类别建模等多个细分方向，需要的同学看文末打包领走

综述

BeyondPixels: A Comprehensive Review of the Evolution of Neural Radiance Fields

神经辐射场演化的全面综述

简述：鉴于NeRF的广泛吸引力，全面检查NeRF的现有研究非常重要。尽管以前的3D渲染调研集中在传统的基于计算机视觉或深度学习的方法，但只有少数讨论了NeRF的潜力。这些调查主要关注NeRF的早期贡献，没有探索它的全部潜力。本文回顾了NeRF的最新进展，并根据架构设计对其进行分类，特别是在新视图合成方面。

NeRF: Neural Radiance Field in 3D Vision, A Comprehensive Review

3D视觉中的神经辐射场，一个全面综述

简述：神经辐射场(NeRF)作为一种新颖的基于隐式场景表示的视图合成方法，已在计算机视觉领域产生巨大影响。鉴于NeRF的普及和研究兴趣，作者认为有必要对过去两年的NeRF论文进行全面综述。作者从架构和应用两个维度对论文进行分类，还介绍了NeRF合成新视图的理论，以及比较了关键NeRF模型的性能和速度。

推理＆训练＆压缩

Hardware Acceleration of Neural Graphics（推理）

神经图形学的硬件加速

简述：论文研究了四个代表性的神经图形学应用，发现当前GPU存在1.5-55倍的性能差距，难以达到理想的渲染性能。作者识别输入编码和多层感知器是性能瓶颈，提出了一个神经图形处理集群架构，通过专用引擎直接加速关键模块，支持各种神经图形学应用。

1. DyLiN: Making Light Field Networks Dynamic

2. RT-NeRF: Real-Time On-Device Neural Radiance Fields Towards Immersive AR/VR Rendering

3. Efficient Neural Radiance Fields for Interactive Free-viewpoint Video

4. R2L： Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis

5. Real-Time Neural Light Field on Mobile Devices

6. Learning Neural Transmittance for Efficient Rendering of Reflectance Fields

7. AUTOMATIC INTEGRATION

8. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks

9. FastNeRF: High-Fidelity Neural Rendering at 200FPS

10. KiloNeRF: Speeding up Neural Radiance Fields with Thousands of Tiny MLPs

11. For Real-time Rendering of Neural Radiance Fields

12. Mixture of Volumetric Primitives for Efficient Neural Rendering

13. Light Field Networks： Neural Scene Representations with Single-Evaluation Rendering

BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis（训练）

为实时视图合成构建神经SDF网格

简述：论文提出了一种方法，用于重建大尺度无界实景的高质量网格，适用于逼真的新视角合成。作者首先优化了一个混合的神经体积-表面场景表示，设计为具有良好的等值面对应场景中的表面。然后，将这个表示“烘焙”成高质量的三角网格，并配备基于球形高斯的简单快速的视点依赖外观模型。最后，优化这个“烘焙”表示以最好地重现捕捉的视点，得到一个模型，可以利用多边形光栅化管线进行实时视图合成。

1. Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction

2. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

3. Radiance Fields without Neural Networks

4. TensoRF: Tensorial Radiance Fields

5. Depth-supervised NeRF: Fewer Views and Faster Training for Free

Variable Bitrate Neural Fields（压缩）

可变比特率神经场

简述：论文提出了一种特征网格压缩字典方法，将其内存消耗降低了高达100倍，并允许多分辨率表示，这对流式处理很有用。作者将字典优化表述为一个向量量化自动解码问题，这使我们可以在一个没有直接监督的空间中端到端学习离散神经表示，拥有动态拓扑结构。

不受约束的图像

1. Ha-NeRF: Hallucinated Neural Radiance Fields in the Wild

2. HDR-Plenoxels: Self-Calibrating High Dynamic Range Radiance Fields

3. NeRF in the Wild： Neural Radiance Fields for Unconstrained Photo Collections

可变形

AligNeRF: High-Fidelity Neural Radiance Fields via Alignment-Aware Training

通过对齐感知训练的高保真神经辐射场

简述：在本文中，作者进行了首个在高分辨率数据上训练NeRF的先导研究，并提出相应的解决方案：1)将多层感知机与卷积层结合，可编码更多邻域信息同时减少参数总数，2)一种新颖的训练策略来解决由移动物体或小的相机标定误差引起的不对齐，3)一个注意高频信息的损失函数。该方法几乎不引入额外的训练/测试成本，实验结果表明，与当前最先进的NeRF模型相比，它可以恢复更多高频细节。

1. DynIBaR: Neural Dynamic Image-Based Rendering

2. D-NeRF: Neural Radiance Fields for Dynamic Scenes

3. Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction

4. PVA: Pixel-aligned Volumetric Avatars

5. Neural Articulated Radiance Field

6. CLA-NeRF: Category-Level Articulated Neural Radiance Field

7. Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies

8. A Higher-Dimensional Representation for Topologically Varying Neural Radiance Fields

9. IBRNet: Learning Multi-View Image-Based Rendering

10. Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes

11. Animatable Neural Radiance Fields from Monocular RGB Videos

12. Neural Actor: Neural Free-view Synthesis of Human Actors with Pose Control

13. TiNeuVox: Fast Dynamic Radiance Fields with Time-Aware Neural Voxels

14. HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular Video

视频＆概括

UV Volumes for Real-time Rendering of Editable Free-view Human Performance（视频）

用于可编辑自由视角人类表演实时渲染的UV体数据集

简述：论文提出了一种新的方法，可以对人类表演者进行实时的自由视角视频渲染。该方法将高频人体外观信息从3D体数据集中分离出来，编码为2D神经纹理栈。使用参数化映射建立人体模型与光滑纹理坐标之间的联系，以实现良好的泛化能力。该方法允许使用更小更浅的网络获得3D信息，同时在2D中捕捉细节。

NeO 360: Neural Fields for Sparse View Synthesis of Outdoor Scenes（概括）

用于户外场景稀疏视图合成的神经场

简述：论文研究了一种新的隐式神经表示方法，称为NeO 360，它可以从单个或很少的RGB图像合成360°全方位视角的户外场景。该方法的核心创新是提出一种混合的基于体素和俯视图的场景表示，它结合了两者的优点，既可以高效表示复杂室外场景，也可以从任意视点进行查询。

1. Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes

2. Space-time Neural Irradiance Fields for Free-Viewpoint Video

3. Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans

4. Neural 3D Video Synthesis from Multi-view Video

5. Dynamic View Synthesis from Dynamic Monocular Video

6. Streaming Radiance Fields for 3D Video Synthesis

7. pixelNeRF: Neural Radiance Fields from One or Few Images

8. Learned Initializations for Optimizing Coordinate-Based Neural Representations

9. pi-GAN: Periodic Implicit Generative Adversarial Networks for 3D-Aware Image Synthesis

10. ShaRF: Shape-conditioned Radiance Fields from a Single View

11. IBRNet: Learning Multi-View Image-Based Rendering

12. CAMPARI: Camera-Aware Decomposed Generative Neural Radiance Fields

13. NeRF-VAE: A Geometry Aware 3D Scene Generative Model

14. Unconstrained Scene Generation with Locally Conditioned Radiance Fields

15. MVSNeRF: Fast Generalizable Radiance Field Reconstruction from Multi-View Stereo

16. Stereo Radiance Fields (SRF): Learning View Synthesis from Sparse Views of Novel Scenes

17. Neural Rays for Occlusion-aware Image-based Rendering

18. Putting NeRF on a Diet: Semantically Consistent Few-Shot View Synthesis

19. MINE: Towards Continuous Depth MPI with NeRF for Novel View Synthesis

20. TöRF: Time-of-Flight Radiance Fields for Dynamic Scene View Synthesis

21. CodeNeRF: Disentangled Neural Radiance Fields for Object Categories

22. StyleNeRF: A Style-based 3D-Aware Generator for High-resolution Image Synthesis

23. Generative Occupancy Fields for 3D Surface-Aware Image Synthesis

24. NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images

25. Point-NeRF: Point-based Neural Radiance Fields

26. SinNeRF: Training Neural Radiance Fields on Complex Scenes from a Single Image

27. Rodin: A Generative Model for Sculpting 3D Digital Avatars Using Diffusion

28. SurfelNeRF: Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes

姿态估计

L2G-NeRF: Local-to-Global Registration for Bundle-Adjusting Neural Radiance Fields

用于捆绑调整神经辐射场的本地到全局注册

简述：论文提出了L2G-NeRF，一种用于捆绑调整神经辐射场的局部到全局注册方法：首先进行基于像素的灵活对齐，然后进行基于帧的受限参数对齐。像素级局部对齐通过深度网络无监督学习，优化光度重建误差。帧级全局对齐使用可微参数估计求解器在像素级对应上找到全局变换。在合成和真实数据集上的实验表明，该的方法在高保真重建和解决大的相机位姿误差方面优于当前最先进的方法。

1. Loc-NeRF: Monte Carlo Localization using Neural Radiance Fields

2. Robust Camera Pose Refinement for Multi-Resolution Hash Encoding

3. iNeRF: Inverting Neural Radiance Fields for Pose Estimation

4. A-NeRF: Surface-free Human 3D Pose Refinement via Neural Rendering

5. NeRF--: Neural Radiance Fields Without Known Camera Parameters

6. iMAP: Implicit Mapping and Positioning in Real-Time

7. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM

8. GNeRF: GAN-based Neural Radiance Field without Posed Camera

9. BARF: Bundle-Adjusting Neural Radiance Fields

10. Self-Calibrating Neural Radiance Fields

组合性

Unsupervised Discovery and Composition of Object Light Fields

无监督发现和组合对象光场

简述：本文提出了一种组成式对象光场表示，可以高效表示和重建物体层面的光信息。该表示将物体建模为光场，并设计了一个组合模块，可以从对象光场重构全局光场。这样既实现了以对象为中心的表示，也大大提升了效率。实验表明，该表示可以实现无监督学习物体表示，同时重建和渲染速度比其他3D方法提升了多个数量级。

1. Learning Object-centric Neural Scattering Functions for Free-viewpoint Relighting and Scene Composition

2. GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields

3. Neural Scene Graphs for Dynamic Scenes

4. Unsupervised Discovery of Object Radiance Fields

5. Learning Object-Compositional Neural Radiance Field for Editable Scene Rendering

6. MoFaNeRF: Morphable Facial Neural Radiance Field

照明

1. KiloNeuS: Implicit Neural Representations with Real-Time Global Illumination

2. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis

3. NeX: Real-time View Synthesis with Neural Basis Expansion

4. NeRFactor: Neural Factorization of Shape and Reflectance Under an Unknown Illumination

5. A Shading-Guided Generative Implicit Model for Shape-Accurate 3D-Aware Image Synthesis

场景标注和理解

1. NeRF-SOS： Any-view Self-supervised Object Segmentation on Complex Real-world Scenes

2. In-Place Scene Labelling and Understanding with Implicit Scene Representation

NeRF编辑

1. Local 3D Editing via 3D Distillation of CLIP Knowledge

2. SINE: Semantic-driven Image-based NeRF Editing with Prior-guided Editing Field

3. General Neural Gauge Fields

4. Editing Conditional Radiance Fields

5. Editable Free-viewpoint Video Using a Layered Neural Representation

6. NeRF-In: Free-Form NeRF Inpainting with RGB-D Priors

7. Unified Implicit Neural Stylization

8. CLIP-NeRF: Text-and-Image Driven Manipulation of Neural Radiance Fields

对象类别建模

1. FiG-NeRF: Figure Ground Neural Radiance Fields for 3D Object Category Modelling

2. NeRF-Tex: Neural Reflectance Field Textures

多尺度

1. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields

2. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields

模型重构

1. NeAT: Learning Neural Implicit Surfaces with Arbitrary Topologies from Multi-view Images

2. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction

3. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction

4. Volume Rendering of Neural Implicit Surfaces

深度估计

NerfingMVS： Guided Optimization of Neural Radiance Fields for Indoor Multi-view Stereo

机器人

1. Differentiable Physics Simulation of Dynamics-Augmented Neural Objects

2. 3D Neural Scene Representations for Visuomotor Control

3. Vision-Only Robot Navigation in a Neural Radiance World

大型场景

1. Switch-NeRF: Learning Scene Decomposition with Mixture of Experts for Large-scale Neural Radiance Fields

2. Block-NeRF: Scalable Large Scene Neural View Synthesis

关注下方《学姐带你玩AI》

回复“NeRF”获取全部118篇论文资料合集

码字不易，欢迎大家点赞评论收藏！