Organizations: School of Computer Science and Technology, Zhejiang Sci-Tech University · School of Biomedical Engineering, Tsinghua University · CardioCloud Medical Technology (Beijing) Co., Ltd. · Lishui University · State Key Laboratory of Blockchain and Data Security, Zhejiang University · College of Artificial Intelligence, Jiaxing University
COLMAP-free 3D Gaussian Splatting (3DGS) bypasses computationally expensive structure-from-motion (SfM) pipelines, yet progressive camera pose tracking remains fundamentally vulnerable to error compounding---early pairwise tracking inaccuracies both corrupt subsequent frame initializations and remain permanently frozen in the scene representation. Rather than relying on heavyweight external neural priors or treating progressive tracking through isolated heuristic fixes, we propose a unified reliability-regulated trajectory optimization framework for progressive COLMAP-free 3DGS. At its core, our framework establishes an intrinsic, self-supervised bidirectional cycle-consistency mechanism that systematically regulates progressive camera trajectory estimation across two complementary temporal horizons: (1) Forward Motion Propagation, where the online reliability signal adaptively gates first-order kinematic warm-starts of rigid motion into upcoming pairwise registrations, supplying informed directional search priors while safely intercepting untrusted transitions; and (2) Retrospective Trajectory Correction, where the same reliability signal dynamically weights relative-pose consistency constraints within a sliding window of neighboring camera poses. By governing both prospective state initialization and retrospective trajectory consolidation through a unified reliability regulator, our self-contained framework resolves progressive drift without external priors or offline preprocessing. Extensive evaluations on Tanks and Temples and CO3D-V2 benchmarks show that our method substantially improves camera trajectory accuracy and novel-view rendering quality, outperforming existing unposed baselines. Code is available at https://github.com/Zijian1026/RRTO-CF3DGS.
Figures & tables
Figure 1: Qualitative comparison on novel-view synthesis and camera trajectory estimation. Our reliability-regulated trajectory optimization prevents local error compounding along the sequence and eliminates trajectory drift relative to CF-3DGS.
Figure 2: Overview of the proposed reliability-regulated trajectory optimization framework. Given an unposed image sequence, our framework establishes an intrinsic bidirectional cycle-consistency signal to systematically regulate progressive camera estimation across two complementary temporal phases: (a) Forward Motion Propagation , where tracking confidence adaptively gates first-order kinematic extrapolation on SE(3) to supply an informed warm-start while intercepting corrupted priors; and (b) Retrospective Trajectory Correction , where camera poses within a sliding window are jointly optimized under fixed Gaussian geometry, constrained by multi-view photometric alignment and confidence-weighted relative pose consistency. Shared cycle reliability ensures that noisy transitions neither corrupt future registrations nor distort local window trajectory refinement.
Figure 3: Qualitative novel-view synthesis results on the Tanks and Temples benchmark. We compare our method against unposed baselines. While existing methods exhibit severe blurring, missing structures, and geometric distortion due to noisy or drifted pose estimates, our approach produces photorealistic renderings with fine-grained textures and sharp edges. Inset red and green boxes highlight key detailed regions.
Scene
Nope-NeRF [ 17 ]
CF-3DGS [ 18 ]
HT-3DGS [ 22 ]
PCR-GS [ 23 ]
Ours
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
Ballroom
24.43
0.71
0.28
24.92
0.83
0.12
25.62
0.85
0.11
25.16
0.83
0.12
28.16
0.91
0.07
Barn
23.00
0.61
0.40
19.26
0.57
0.38
20.33
0.56
0.35
19.00
0.53
0.46
22.01
0.67
0.27
Church
23.01
0.65
0.37
26.83
0.87
0.14
27.80
0.89
0.12
27.31
0.89
0.13
27.53
0.89
0.12
Family
24.59
0.71
0.35
25.03
0.85
0.16
27.80
0.89
0.12
27.15
0.90
0.12
28.31
0.91
0.11
Francis
23.01
0.66
0.43
23.90
0.74
0.27
24.31
0.72
0.26
24.77
0.73
0.26
25.96
0.77
0.25
Table 1: Quantitative novel view synthesis results on the Tanks and Temples benchmark with stride 5. Held-out test views are evaluated using PSNR ↑ , SSIM ↑ , and LPIPS ↓ . Best results are bolded.
Figure 4: Qualitative comparison on challenging 360-degree object-centric sequences from CO3D-V2. On the displayed sequences with wide baselines and continuous rotational trajectories, competing methods exhibit trajectory deviations and rendering artifacts such as floaters and blurriness. Our reliability-regulated framework produces more accurate trajectory estimates and higher-fidelity Gaussian reconstructions.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
CF-3DGS [ 18 ]
26.11
0.782
0.270
HT-3DGS [ 22 ]
26.59
0.800
0.260
PCR-GS [ 23 ]
26.16
0.785
0.272
Ours
27.71
0.819
0.239
Table 2: Average novel view synthesis results on the CO3D-V2 dataset. Test views are evaluated using PSNR ↑ , SSIM ↑ , and LPIPS ↓ . Best results are bolded.
Scene
Nope-NeRF [ 17 ]
CF-3DGS [ 18 ]
HT-3DGS [ 22 ]
PCR-GS [ 23 ]
Ours
RPEt
RPEr
ATE
RPEt
RPEr
ATE
RPEt
RPEr
ATE
RPEt
RPEr
ATE
RPEt
RPEr
ATE
Ballroom
0.127
0.247
0.002
0.279
0.048
0.006
0.279
0.047
0.006
0.295
0.059
0.006
0.122
0.026
0.002
Barn
1.343
0.403
0.033
2.409
0.585
0.065
2.252
0.574
0.057
3.379
1.074
0.118
0.576
0.314
0.011
Church
0.490
0.127
0.044
0.081
0.095
0.004
0.081
0.095
0.004
0.070
0.092
0.005
0.053
0.087
0.003
Family
0.119
0.052
0.003
0.775
0.294
0.011
0.203
0.118
0.004
0.483
0.187
0.006
0.455
0.181
0.006
Francis
2.941
1.515
0.079
0.814
0.916
0.021
0.733
0.878
0.016
0.714
0.816
0.015
0.416
0.747
0.015
Table 3: Quantitative camera pose estimation results on the Tanks and Temples benchmark with stride 5. Trajectories are evaluated using RPEt↓ ( ×102 ), RPEr↓ ( ∘ ), and ATE ↓ . Best results are bolded.
Method
RPEt↓
RPEr↓
ATE ↓
CF-3DGS [ 18 ]
0.579
0.548
0.023
HT-3DGS [ 22 ]
0.545
0.511
0.022
PCR-GS [ 23 ]
0.658
0.842
0.032
Ours
0.352
0.382
0.014
Table 4: Average camera pose estimation results on the CO3D-V2 dataset. Trajectories are evaluated using RPEt↓ ( ×102 ), RPEr↓ ( ∘ ), and ATE ↓ . Best results are bolded.
Figure 5: Reliability verification of bidirectional cycle-consistency confidence. We visualize the temporal trajectory error against the estimated cycle-consistency confidence ci across representative sequences from Tanks and Temples ( A : Family , B : Horse ) and CO3D-V2 ( C : Apple , D : Book ). For each sequence, the upper subplot shows the frame-wise bidirectional confidence ci , while the lower subplot displays the corresponding translational error and rotational discrepancy.
Variant
Controls
Loss Weight
Rendering quality
Pose accuracy
MP
Gate
WJ
PSNR ↑
SSIM ↑
LPIPS ↓
RPEt↓
RPEr↓
ATE↓
CF-3DGS (sequential)
×
–
×
–
21.93
0.709
0.250
1.327
0.695
0.024
WJ, photometric only
×
–
✓
–
22.56
0.724
0.247
1.047
0.623
0.022
WJ + Uniform loss weight
×
–
✓
Uniform
22.74
0.726
0.242
0.998
0.618
0.021
WJ + F–B loss weight
×
–
✓
F–B
22.97
0.729
0.237
0.920
0.610
0.019
Ungated MP
✓
×
×
–
23.58
0.752
0.210
0.714
0.476
0.016
Table 5: Ablation study and reliability calibration on Tanks and Temples. We evaluate the impact of motion-guided initialization (MP), confidence gating (Gate), window joint refinement (WJ), and forward–backward (F–B) consistency weighting. RPEt is reported in ×102 and RPEr in degrees.
Pose-free 3D Gaussian Splatting (3DGS) aims to bypass cumbersome Structure-from-Motion (SfM) by directly reconstructing scenes from unposed image sequences. In this work, we present CrossFit-GS, a deferred-commitment progressive framework that mitigates cumulative trajectory drift and breaks the self-reinforcing loop of distorted-map error propagation. Specifically, we partition the input sequence into interleaved subsets and maintain two complementary submaps. Each candidate pose is cross-refined against a frozen opposite submap that has never been optimized using the target frame, thereby providing an independent geometric reference for pose estimation. Checkerboard validation and spatial coverage safeguards further reject unreliable updates, and only verified poses are committed to their corresponding owner submap. Once the trajectory has been refined, the intermediate submaps are discarded and the final global representation is trained from scratch using the fixed refined poses. Extensive evaluations across multiple benchmarks demonstrate that CrossFit-GS substantially reduces cumulative trajectory drift and consistently improves both camera pose estimation and novel-view synthesis over existing methods.
Zijian Wu, Jinliang Wang, Mingfeng Jiang +3
Zhejiang Sci-Tech University · Tsinghua University
Feature extraction, matching, structure from motion (SfM), and novel view synthesis (NVS) have traditionally been treated as separate problems with independent optimization objectives. We present GloSplat, a framework that performs joint pose-appearance optimization during 3D Gaussian Splatting training. Unlike prior joint optimization methods (BARF, NeRF--, 3RGS) that rely purely on photometric gradients for pose refinement, GloSplat preserves explicit SfM feature tracks as first-class entities throughout training: track 3D points are maintained as separate optimizable parameters from Gaussian primitives, providing persistent geometric anchors via a reprojection loss that operates alongside photometric supervision. This architectural choice prevents early-stage pose drift while enabling fine-grained refinement -- a capability absent in photometric-only approaches. We introduce two pipeline variants: (1) GloSplat-F, a COLMAP-free variant using retrieval-based pair selection for efficient reconstruction, and (2) GloSplat-A, an exhaustive matching variant for maximum quality. Both employ global SfM initialization followed by joint photometric-geometric optimization during 3DGS training. Experiments demonstrate that GloSplat-F achieves state-of-the-art among COLMAP-free methods while GloSplat-A surpasses all COLMAP-based baselines.
Tianyu Xiong, Rui Li, Linjie Li +1
Department of Computer Science, Northwestern Polytechnical University, Shanxi, China · CEMSE, KAUST, Thuwal, Jeddah
Pose-free feed-forward 3D Gaussian Splatting (3DGS) reconstructs a scene from sparse, unposed images in a single network pass, removing the need for camera calibration and per-scene optimization. However, camera estimation errors propagate into the predicted Gaussians and compound the geometric and photometric inaccuracies of single-pass prediction. To correct these errors, we introduce DeltaSplat, a lightweight Gaussian refinement module for pose-free feed-forward 3DGS. It iteratively renders the current Gaussians at the input context views and predicts per-Gaussian updates from the resulting residuals. A 2D residual alone, however, underdetermines the 3D correction. DeltaSplat therefore conditions each update on per-pixel Plücker rays and rendered depth as a soft geometric prior. A dual-branch convolutional mixer efficiently encodes these inputs, and per-attribute heads decode the fused features into position, opacity, and color updates. The module adds only ~2.2% parameters to the backbone and remains fully feed-forward at inference. On DL3DV, DeltaSplat reaches 26.64 dB PSNR in the pose-free setting, improving its state-of-the-art backbone by 1.75 dB and surpassing even baselines supplied with ground-truth cameras; consistent gains hold across 6-24 views and all camera regimes.
Chanung Park, Seunghyeon Song, Joo Chan Lee +2
Department of Electrical and Computer Engineering, Sungkyunkwan University · Department of Artificial Intelligence, Sungkyunkwan University · Department of Artificial Intelligence, Yonsei University