A Survey on Pose-Free Neural Radiance Fields and 3D Gaussian Splatting
A comprehensive review of pose-free neural rendering and 3D reconstruction, covering NeRF and 3DGS methods with only noisy pose estimation or without camera pose priors.
Dongbo Shi1, Lubin Fan2, Bojian Wu2,
Shen Cao2, Jinhui Guo2, Ligang Liu1, Renjie Chen1
1University of Science and Technology of China
2Independent Researcher
1. Overview
Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have significantly advanced 3D scene reconstruction and novel view synthesis in recent years.
Despite continuous improvements in accuracy, quality, and application scope, these methods fundamentally rely on a critical assumption: precise camera poses for all input images.
The reconstruction pipeline typically employ multi-view geometry techniques such as Structure-from-Motion (SfM) or Simultaneous Localization and Mapping (SLAM) to estimate these camera poses.
However, in many real-world cases, SfM/SLAM may fail due to weak textures, large viewpoint changes, or non-sequential inputs.
This shifts the focus to the joint estimation of scene structure and camera motion within a unified framework.
Understanding and addressing this challenge is essential to unlock the full potential of neural 3D reconstruction in unconstrained real-world environments.
To address these challenges of 3D scene reconstruction without reliable camera parameters, recent research has attracted growing interest in adapting and extending neural rendering paradigms (particularly NeRF and 3DGS) to perform pose-free or pose-robust reconstruction.
This survey systematically reviews these pose-free approaches into following categories:
- Base model enhancements
- Training strategy refinements
- Novel priors incorporation
The survey also introduces evaluation methodologies through standardized datasets and metrics, and benchmarks representative methods.
2. Method Taxonomy
Neural Radiance Fields (NeRF)
2.1 Base Model Enhancement
Network Architecture: GNeRF, SiNeRF, GARF, DBARF
Robust NeRF: RM-NeRF, Robust-RMNeRF, DiGARR, I-DACS
2.2 Training Strategy Refinement
Coarse-to-Fine: BARF, CBARF
Local-to-Global: LU-NeRF, CF-NeRF, CT-NeRF
Sampling Strategies: NeRFmm, iNeRF, CoPoNeRF, TD-NeRF, FDC-NeRF
Pose Estimators: Camp, SaNeRF, NeRFtrinsic Four, IR-NeRF, PoseProbesGenObj, CAD-NeRF
2.3 Novel Prior Incorporation
Depth-based: Bid-NeRF, NoPe-NeRF, AltNeRF, RGBD-NeSR
Feature-based: SC-NeRF, L2G-NeRF, InvWarp-NeRF, FlowCam, RoGUENeRF
LiDAR-based: LiDeNeRF, GeoNLF
Multiview Geometry: DA-NeRF, RPE-BARF, NoPe-NeRF++
Surface Geometry: SG-NeRF, RSG-NeRF, PoRF
2.4 Applications
In-the-wild: SAMURAI, NeROIC, SHINOBI, UP-NeRF
Motion Images: BAD-NeRF, BeNeRF, EBAD-NeRF, USB-NeRF, RS-NeRF
Sparse View: RegNeRF, SPARF, DaRF, SN2eRF, GC-NeRF, TrackNeRF, Just Flip
Large-scale: LocalRF, UC-NeRF, DGNR
Dynamic Scene: BASED, DynaMoN
3D Gaussian Splatting (3DGS)
2.1 Base Model Enhancement
No existing works fall into this category.
2.2 Training Strategy Refinement
Progressive Optimization: CF-3DGS, GGRt
Grouping Strategies: EasySplat, SFGS, KeyGS, Rob-GS
2.3 Novel Prior Incorporation
Depth-based: PF3Plat, LG-3DGS, ZeroGS, SelfSplat
Feature-based: HybridBA-3DGS, TrackGS, PCR-GS, 3R-GS
2.4 Applications
Motion Images: SC4D-3DGS, DreamScene4D, MotionGS, Deblur-GS, EF-3DGS, IncEventGS
Sparse View: InstantSplat, GaussianScenes, MetaSplats, GBR, GeSplat, NoPoSplat, SPFSplat, COGS
Medical Surgery: Free-SurGS, Free-DyGS
3. Benchmarks & Datasets
The following datasets are widely used for evaluating pose-free NeRF and 3DGS reconstruction.
Their characteristics differ in pose availability, scene type, trajectory patterns, and scale.
| Dataset |
Camera Poses |
Synthetic |
Scene Type |
Trajectory Type |
#Views |
#Scenes |
| NeRF Synthetic |
Ground-truth |
Yes |
Indoor |
Object-centric |
~100 |
8 |
| LLFF |
Known |
No |
Indoor |
Forward-facing |
20–60 |
8 |
| DTU |
Known |
No |
Indoor |
Object-centric |
49–64 |
124 |
| Replica |
Known |
Yes |
Indoor |
Complex trajectory |
~2K |
18 |
| Tanks and Temples |
Known |
No |
Outdoor |
Complex trajectory |
~200 |
8 |
| RealEstate10K |
Unknown |
No |
Indoor |
Complex trajectory |
Video |
~70K |
| CO3D V2 |
Estimated |
No |
Mixed |
Object-centric |
~200 |
~40K |
| Static Hikes |
Estimated |
No |
Outdoor |
Complex trajectory |
~10K |
12 |
Synthetic datasets provide accurate pose annotations, while real‑world datasets capture
diverse illumination, large motion baselines, and complex geometry—making them essential for benchmarking
robust pose‑free reconstruction.
Dataset Gallery
Representative scenes from major datasets can be visualized below.
Figure: Dataset gallery
4. Resources
We maintain a comprehensive Awesome Pose-Free NeRF & 3DGS list on GitHub,
covering papers, datasets, benchmarks, and open-source implementations used in this survey.
🧪 Datasets for Pose-Free Reconstruction
Includes LLFF, DTU, CO3D, Replica, RealEstate10K, Tanks and Temples, etc.
→ View Dataset Links
5. Citation
@article{posefree_survey_2025,
title={A Survey on Pose-Free Neural Radiance Fields and 3D Gaussian Splatting},
author={Dongbo Shi,Lubin Fan,Bojian Wu,Shen Cao,Jinhui Guo,Ligang Liu,Renjie Chen},
year={2025}
}