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.
Figures & tables
Figure 1: Accuracy vs. Speed. Average PSNR on MipNeRF360 vs. runtime for 1000 images (Courthouse scene). GloSplat-F achieves 13.3 × speedup over GPU-accelerated COLMAP+3DGS while improving PSNR by +0.38 dB. GloSplat-A achieves the highest accuracy (28.86 dB), surpassing all baselines. Our joint pose-appearance optimization enables both variants to occupy the Pareto frontier. (All methods benchmarked on the same GPU; see Appendix C for details.)
Figure 2: GloSplat Pipeline. Given unposed input images, local correspondences are extracted (frozen preprocessing): XFeat+LightGlue with retrieval-based pairs (GloSplat-F) or SIFT with exhaustive matching (GloSplat-A). Global SfM simultaneously estimates all camera poses through rotation averaging, positioning, and bundle adjustment, providing robust initialization. Joint 3DGS training (our core contribution) then continuously refines poses through a reprojection-based BA loss while optimizing Gaussian primitives, enabling combined photometric-geometric supervision that prevents drift and improves reconstruction quality.
MipNeRF360
Tanks and Temples
CO3Dv2
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
3DGS †
27.39
0.815
0.185
24.85
0.851
0.155
32.58
0.938
0.095
MCMC †
27.91
0.836
0.154
25.76
0.867
0.139
33.21
0.941
0.097
MCMC
22.19
0.548
0.282
21.42
0.679
0.278
25.71
0.712
0.201
CF-3DGS
12.38
0.234
0.719
12.19
0.391
0.608
20.18
0.611
0.435
HT-3DGS
14.79
0.380
0.669
13.83
0.451
0.585
28.28
0.833
0.230
Table 1: Comparison with state-of-the-art methods on novel view synthesis. † indicates COLMAP initialization. Best COLMAP-free results in bold . GloSplat-F (retrieval-based) achieves SOTA among COLMAP-free methods.
Figure 3: Runtime Comparison on Courthouse Scene. End-to-end reconstruction time (in seconds) as a function of the number of input images. All methods use the same GPU (RTX PRO 6000); COLMAP is compiled with CUDA and uses GPU acceleration for feature extraction/matching. GloSplat-F achieves 13.3 × speedup over COLMAP at 1000 images due to retrieval-based pair selection and parallel global SfM. VGGT-X is faster at smaller scales but GloSplat-F surpasses it at 750+ images due to better asymptotic scaling.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
3DGS
27.48
0.815
0.216
AbsGS
27.52
0.820
0.198
PixelGS
27.62
0.824
0.189
MiniSplatting-D
27.57
0.832
0.176
TamingGS
27.96
0.822
0.207
3DGS-MCMC
28.01
0.835
0.186
Table 2: Comparison with COLMAP-based methods on MipNeRF360. All baselines use COLMAP poses. GloSplat-A achieves SOTA among all methods. Best in bold , second best underlined .
COLMAP
3RGS
GloSplat-F (Ours)
Scene
R(°) ↓
ATE(m) ↓
PSNR ↑
R(°) ↓
ATE(m) ↓
PSNR ↑
R(°) ↓
ATE(m) ↓
PSNR ↑
0079_00
3.55
0.014
30.78
2.45
0.014
32.58
2.12
0.011
33.42
0301_00
133.83
0.169
23.63
9.30
0.009
30.11
7.82
0.007
31.24
0418_00
5.03
0.012
29.03
4.34
0.012
31.62
3.78
0.010
32.35
Table 3: Comparison with COLMAP and 3RGS on ScanNet. Best in bold , second best underlined .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Qualitative Comparison on MipNeRF360. We compare novel view synthesis results from GloSplat-A, GloSplat-F, VGGT-X, and Improved-GS against ground truth. GloSplat-A achieves significantly higher PSNR and lower LPIPS across all scenes. On Bonsai , GloSplat-A outperforms VGGT-X by +8.57 dB PSNR . On Flowers , our method shows +6.76 dB PSNR and 46% lower LPIPS (0.147 vs 0.273) over VGGT-X. On Garden , GloSplat-A achieves +5.69 dB over VGGT-X. Best viewed zoomed in.
Scene
PSNR ↑
SSIM ↑
LPIPS ↓
Apple
35.30
0.952
0.063
Bench
32.33
0.931
0.089
Hydrant
29.33
0.920
0.071
Skateboard
33.35
0.943
0.116
Teddybear
33.24
0.932
0.102
Average
32.71
0.936
0.088
Appendix
Table 4: Per-scene results on CO3Dv2 dataset.
Scene
PSNR ↑
SSIM ↑
LPIPS ↓
Bicycle
25.98
0.802
0.145
Bonsai
33.16
0.952
0.114
Counter
29.60
0.921
0.131
Flowers
22.93
0.688
0.244
Garden
27.64
0.866
0.084
Kitchen
32.99
0.942
0.085
Appendix
Table 5: Per-scene results on MipNeRF360 dataset.
Scene
PSNR ↑
SSIM ↑
LPIPS ↓
Barn
26.43
0.842
0.168
Caterpillar
24.89
0.851
0.162
Ignatius
27.15
0.912
0.121
Truck
25.31
0.867
0.148
Train
25.32
0.873
0.156
Average
25.82
0.869
0.151
Appendix
Table 6: Per-scene results on Tanks and Temples dataset.
Method
250 imgs
500 imgs
750 imgs
1000 imgs
COLMAP
644.6
1165.0
2226.3
4349.2
3RGS (MASt3R)
312.2
638.0
978.5
1308.2
GloSplat-A
222.2
408.2
855.7
1371.3
VGGT-X
53.9
113.3
249.6
398.9
GloSplat-F
100.2
191.9
250.0
327.8
Speedup vs COLMAP
Appendix
Table 7: Runtime comparison (seconds) on Courthouse scene with varying image counts. All methods run on the same hardware (RTX PRO 6000 GPU); COLMAP uses GPU-accelerated SIFT and matching. GloSplat-F achieves superior scaling due to O(n) pair selection and parallel global SfM.
Configuration
Training Time
Overhead
MCMC 3DGS (vanilla)
∼ 31 min
–
MCMC 3DGS + Joint Opt (Ours)
∼ 32 min
+3%
Appendix
Table 8: Training time comparison between vanilla MCMC 3DGS and MCMC 3DGS with joint optimization (30k iterations, 200 images). Joint optimization adds minimal overhead while providing +0.81 dB improvement.
Configuration
PSNR ↑
Δ
SSIM ↑
LPIPS ↓
Reference Configurations
GloSplat-A (exhaustive matching)
28.86
+1.09
0.862
0.139
GloSplat-F (retrieval matching)
27.77
–
0.818
0.164
MCMC † (COLMAP, frozen poses)
27.91
+0.14
0.836
0.154
(A) Joint Optimization — removing from GloSplat-F
Frozen poses after Global SfM
19.18
–8.59
0.463
0.472
Appendix
Table 9: Ablation study on MipNeRF360. We show reference configurations (top) and ablations organized by component. Δ shows change from GloSplat-F. Best in bold ; “–” indicates failure.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
GloSplat-A (separate tracks)
28.86
0.862
0.139
GloSplat-A (merged tracks)
28.64
0.858
0.144
Δ
–0.22
–0.004
+0.005
GloSplat-F (separate tracks)
27.77
0.818
0.164
GloSplat-F (merged tracks)
27.58
0.813
0.168
Δ
–0.19
–0.005
+0.004
Appendix
Table 10: Ablation: Separating track points from Gaussian means. Merging tracks with Gaussians degrades all metrics for both variants, validating our architectural choice.