Large-scale scene reconstruction is a critical foundational technology in robotic autonomous systems such as 3D mapping and autonomous driving. In recent years, 3D Gaussian Splatting (3DGS) has demonstrated remarkable advantages in both visual quality and computational efficiency, making it a promising representation for large-scale scene reconstruction. However, it still faces challenges in large-scale scenes, including excessive memory consumption and uneven viewpoint coverage caused by UAV acquisition, limiting its real-world applications. To address this, we propose VDGS, a novel 3DGS framework that incorporates camera distribution into scene modeling. VDGS introduces visibility-driven statistics for scene anchors to quantify supervision strength. These statistics are further leveraged for scene partitioning and for gradient compensation in under-optimized regions, thereby promoting balanced optimization across different regions. Extensive experiments on multiple large-scale aerial scene datasets demonstrate that, under imbalanced viewpoint distributions, VDGS consistently outperforms existing methods, while maintaining competitive performance in scenarios with more uniform view distributions.
Figures & tables
Fig. 1: Top : 3DGS adopts a view-driven optimization scheme. The colored arrows associated with each Gaussian indicate gradient contributions from cameras of corresponding colors. Within one optimization cycle iterating over four camera views once, the red Gaussian receives three updates, whereas the blue Gaussian is updated only once. Bottom : Visualization of camera distribution and point visibility in the Residence scene [ 19 ] . The color of each point ranges from cool to warm, indicating low to high visibility. Point visibility serves as a proxy for the degree of optimization.
Fig. 2: Overview of the proposed VDGS framework. (a) Visibility-Driven Data Partitioning. The scene is partitioned into blocks using a visibility-aware clustering strategy, where clustering and partitioning are performed on planar primitives extracted from an initially trained coarse model. (b) Re-rendering Error Based Camera Assignment. A camera view is assigned to a block when the rendered area of that block in the view exceeds a predefined threshold. (c) Visibility-Guided Anchor Generation. During training, gradients of low-visibility anchors are accumulated into a global buffer and later used for gradient compensation. This reinforcement alleviates weak supervision and promotes anchor densification in under-observed regions. (d) Block Merge. After training, all blocks are merged according to visibility criteria to produce the final rendering result.
Fig. 3: Partition pipeline. First, feature vectors are constructed for anchors using planar x – y coordinates and visibility statistics, and anchors are clustered in the resulting feature space. Next, connectivity filtering is performed within each cluster to retain the largest connected component, based on which an α -shape defines the block boundary. Finally, the block boundaries are further expanded to obtain the final block-level partitions.
Method
Residence
Sci-Art
Rubble
Building
Block-All
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Switch-NeRF [ 27 ]
22.57
0.654
0.352
26.51
0.795
0.271
24.31
0.562
0.478
21.54
0.579
0.397
–
–
–
3DGS [ 9 ]
21.94
0.764
0.279
21.85
0.787
0.311
25.20
0.757
0.318
22.04
0.728
0.332
27.36
0.818
0.237
Scaffold-GS [ 15 ]
22.00
0.761
0.286
22.56
0.796
0.302
24.83
0.721
0.353
22.42
0.719
0.335
27.13
0.868
0.210
DOGS [ 12 ]
21.94
0.740
0.244
24.42
0.804
0.219
25.78
0.765
0.257
22.73
0.759
0.214
28.58
0.847
0.219
Octree-GS [ 16 ]
22.29
0.762
0.288
23.38
0.828
0.240
25.34
0.763
0.299
23.66
0.776
0.267
26.41
0.814
0.282
TABLE I: Quantitative comparison of our VDGS with prior methods across five large-scale aerial scenes. We report PSNR ( ↑ ), SSIM ( ↑ ), and LPIPS ( ↓ ) on test views. The best and second-best scores are highlighted.
Fig. 4: Qualitative rendering comparisons on the Mill19 [ 8 ] and UrbanScene3D [ 19 ] datasets. Zoomed-in views of the regions highlighted by red boxes are shown for detailed inspection.
Fig. 5: Camera and point visibility distributions for the Building and Sci-Art scenes. Camera frustums are red; warmer point colors denote higher visibility. Building shows relatively uniform camera placement and visibility, whereas Sci-Art exhibits more uneven coverage and less consistent visibility.
Training Strategy
PSNR ↑
SSIM ↑
LPIPS ↓
(a) Baseline
22.64
0.801
0.233
(b) w/o Blocking
23.39
0.827
0.209
(c) w/o Anchor Generation
23.56
0.833
0.202
(d) Full
24.01
0.845
0.188
TABLE II: Ablation study on different training strategies evaluated on the Residence scene. We report PSNR ( ↑ ), SSIM ( ↑ ), and LPIPS ( ↓ ).
#Blocks
PSNR ↑
SSIM ↑
LPIPS ↓
1
23.39
0.827
0.209
2
23.63
0.835
0.196
4
24.01
0.845
0.188
8
24.04
0.846
0.187
16
24.05
0.846
0.187
TABLE III: Ablation study on the number of blocks evaluated on the Residence scene. We report PSNR ( ↑ ), SSIM ( ↑ ), and LPIPS ( ↓ ).
Parameter / Value
PSNR ↑
SSIM ↑
LPIPS ↓
(a) Visibility weight λvis
λvis=0 (spatial only)
23.51
0.8353
0.1988
λvis=0.5
23.99
0.8433
0.1918
λvis=1.0 (default)
24.03
0.8453
0.1884
λvis=1.3
23.80
0.8381
0.1928
(b) Low-visibility threshold Tlow
TABLE IV: Ablation on key hyperparameters evaluated on the Residence scene. Each parameter is varied independently while keeping the others at their defaults ( λvis=1.0 , Tlow=0.20 , βg=2.0 , γ=0.90 ). PSNR ( ↑ ), SSIM ( ↑ ), LPIPS ( ↓ ).