Lightweight 3D building models support urban spatial analysis, but their production often requires dense scene reconstruction, building extraction, and mesh simplification. Although 3D Gaussian Splatting (3DGS) is more efficient than conventional multi-view stereo, existing large-scale methods lack semantic selection and parallel instance-level training. We present SemLite-GS, a building-centric framework that reconstructs lightweight building meshes from large-scale UAV oblique imagery. First, an orthographic reference map establishes globally consistent building identities, which are propagated to oblique views to form camera-instance pairs and regions of interest. Second, semantic-selective parallel PGSR training jointly processes multiple pairs within relevant image regions while updating only target Gaussians; contextual Gaussians remain fixed to preserve visibility. Third, architectural planes extracted from rendered depth, opacity, and normal maps guide signed distance field optimization, producing meshes with adaptive polygon allocation and well-defined structural boundaries. Experiments on multiple large-scale UAV datasets demonstrate that our SemLite-GS substantially reduces building reconstruction time compared with representative full-scene 3DGS methods. With comparable polygon budgets, SemLite-GS achieves superior preservation of dominant planar structures and sharp architectural transitions compared with conventional dense reconstruction followed by mesh decimation. The source code is available at https://github.com/IRIS-LAB-whu/SemLite-GS.
Figures & tables
Figure 1: Qualitative comparison of building meshes generated by SemLite-GS and Metashape ( LLC, 2022 ) . The upper panels show individually identified building meshes from SemLite-GS and a continuous scene mesh from Metashape. The lower panels show the building marked by the red boxes from a different oblique viewpoint, highlighting differences in planar triangulation and building-edge regularity.
Figure 2: Overview of SemLite-GS. A FastGS initializes a shared scene-level Gaussian representation from registered UAV images and their camera parameters. Orthographically anchored instance propagation establishes globally consistent building identities and constructs a camera–instance observation graph with verified masks and ROIs. Semantic-selective parallel PGSR refines multiple building instances concurrently within target-specific Gaussian and image domains, retaining contextual Gaussians for visibility-consistent rendering while updating only the corresponding targets. The refined Gaussians provide multi-view RGB, depth, opacity, and normal observations. Finite architectural planes estimated from the masked geometric observations constrain Gaussian-initialized SDFs and guide plane-aware surface consolidation and triangulation to produce lightweight building meshes. Arrow colors distinguish scene and data flow, semantic anchoring, selective training, and geometric constraints; the dashed arrow indicates the reuse of finite plane supports during mesh extraction.
Figure 3: Workflow of orthographically anchored instance propagation. (a) Orthographic semantic anchoring: the FastGS initializes a shared Gaussian representation for DOM rendering. SAM3 predictions from overlapping DOM tiles are merged into a global instance map, and building identities are assigned to Gaussians through local support voting and verification of height and connectivity. (b) Visibility-consistent propagation: semantic rendering accounts for full-scene occlusion, with all Gaussians determining transmittance and target-labeled Gaussians contributing instance support. Visibility verification and mask refinement produce camera–instance observations, which are retained through visible-area filtering to construct the observation graph. Each edge stores a verified mask Mc,q and its associated ROI Ωc,q .
Figure 4: Workflow of semantic-selective parallel PGSR. Each iteration selects one valid camera observation for each sampled building. Pair-conditioned compaction packs the active Gaussian indices, and ROI-constrained rasterization processes the resulting pair–tile tasks concurrently to produce pair-local rendering buffers. Target-mask supervision and frozen-context protection guide selective updates through parallel backpropagation and target-gradient writeback. Orange, teal, and purple indicate distinct target instances, while gray denotes contextual Gaussians within each pair. Gaussian numbers identify shared scene indices, and queue entries use the format pair:tile. The return arrow indicates the next optimization iteration. Loss weights are omitted for clarity.
Figure 5: Pipeline of plane-constrained lightweight mesh reconstruction. Multi-view depth, opacity, and normal observations within verified target masks are fused to estimate finite architectural planes. These planes regularize a Gaussian-initialized SDF and guide the consolidation of planar surface regions during extraction. Planar and residual patches are triangulated with shared boundaries under a prescribed polygon budget. The dashed arrow indicates the reuse of finite plane supports for plane association.
Dataset
Scene
# Images
Resolution
Target instances
LiDAR
GauScene
CUHK_LOWER
670
5472×3648
18
Yes
CUHK_UPPER
715
5472×3648
22
Yes
SZIT
1,215
5472×3648
26
Yes
Skylume
building
3,646
1563×1039
24
Yes
tec_school
2,394
6260×4165
11
Yes
Tianjin_Xiqing
Tianjin_Xiqing
305
3942×2930
34
No
Table 1: Overview of the selected scenes. Resolution denotes the dimensions (width × height in pixels) of the highest-pixel-count image available for each scene. Target instances indicates the number of building instances in each scene.
Figure 6: Instance-level semantic selectivity on the Tianjin_Xiqing scene. (a) Orthophoto overlaid with automatically generated building-instance masks. (b) Gaussian rendering of the corresponding building instances after selective PGSR training by SemLite-GS. (c) Full-scene Gaussian reconstruction by CityGS-X. (d) Full-scene Gaussian reconstruction by CityGaussianV2. The panels are shown from aligned overhead viewpoints.
Figure 7: Qualitative comparison of lightweight building meshes produced by seven pipelines across six scenes. Columns show CUHK_LOWER, CUHK_UPPER, SZIT, building, tec_school, and Tianjin_Xiqing. The upper rows present scene overviews; dashed boxes identify the buildings shown from oblique viewpoints in the lower rows. Corresponding building instances use consistent colors across methods.
Dataset
Scene
Valid instances
Pipeline
CD (m) ↓
Avg. vertices ↓
Avg. faces ↓
GauScene
CUHK_LOWER
18
CityGS-X + QEM
2.305
282
1000
CityGaussianV2 + QEM
1.559
384
1000
SemLite-GS (TSDF) + QEM
1.528
407
1000
PG-SAG + QEM
1.373
563
1000
PG-SAG (mask) + QEM
1.031
578
1000
Metashape + QEM
1.154
582
1000
Table 2: Scene-level comparison of geometric accuracy and mesh complexity for lightweight building meshes across five scenes with LiDAR references. Metashape+QEM uses building meshes extracted from the Metashape scene reconstruction and simplified using QEM. PG-SAG (mask)+QEM replaces PG-SAG’s original building masks with those generated by SemLite-GS. QEM-based meshes are simplified to 1,000 faces per building in four scenes and 400 faces in tec_school, whereas SemLite-GS determines the face count adaptively for each building. CD, vertex count, and face count are computed per building and averaged within each scene. The Overall row reports instance-weighted means over all 101 evaluated buildings.
CD (m) ↓
Dataset
Scene
Valid instances
SemLite-GS
PG-SAG
PG-SAG (mask)
CityGaussianV2
CityGS-X
Metashape
GauScene
CUHK_LOWER
18
0.833
1.306
0.952
1.214
1.455
1.140
CUHK_UPPER
22
0.807
0.972
0.902
1.033
1.904
1.165
SZIT
26
0.972
1.344
1.022
1.241
1.633
1.574
Skylume
building
24
0.481
1.043
0.693
0.705
1.148
0.611
tec_school
11
0.445
0.657
0.484
0.601
0.984
0.681
Table 3: Geometric accuracy of dense building surface meshes before QEM simplification. Gaussian-based methods extract meshes by TSDF fusion of rendered multi-view depth maps, whereas Metashape produces meshes through its photogrammetric reconstruction pipeline. PG-SAG (mask) denotes the mask-substitution variant described above. CD is computed per building and averaged within each scene; the overall row averages all 101 evaluated buildings.
Pipeline
Preprocessing (h) ↓
Training (h) ↓
Mesh reconstruction (h) ↓
Total time (h) ↓
CityGS-X + QEM
0.40
12.19
1.49
14.08
CityGaussianV2 + QEM
0.43
72.38
1.06
73.87
PG-SAG + QEM
1.07
9.29
0.64
11.00
Metashape + QEM
–
–
3.57
3.57
SemLite-GS
0.64
2.31
0.53
3.48
Table 4: Cumulative wall-clock runtime for lightweight building-mesh reconstruction across five scenes. Stage-level times are reported for the 3DGS-based pipelines. For Metashape+QEM, only the total software execution time is reported, covering registered-image and camera-pose import, dense-scene mesh reconstruction, instance extraction, and QEM simplification. Manual target selection is excluded.
Figure 8: Qualitative comparison of lightweight building meshes reconstructed by SemLite-GS, PolyFit, KSR, and Compod across six scenes. Columns correspond to CUHK_LOWER, CUHK_UPPER, SZIT, building, tec_school, and Tianjin_Xiqing. The upper four rows show scene overviews; the fractions indicate successfully reconstructed buildings relative to the target count. The lower four rows show enlarged views of buildings marked by dashed boxes. “Timeout (900 s)” indicates that PolyFit did not produce a mesh for the selected building within the allotted time. Colors denote corresponding building instances across methods. Tianjin_Xiqing is included for qualitative assessment only because LiDAR reference data are unavailable.
Dataset
Scene
Valid instances
Method
CD (m) ↓
Avg. vertices ↓
Avg. faces ↓
GauScene
CUHK_LOWER
18
PolyFit 10/18
3.380
204
208
KSR
5.729
441
792
Compod
2.670
414
860
SemLite-GS
1.026
546
998
CUHK_UPPER
22
PolyFit 5/22
3.906
115
101
KSR
4.215
414
772
Table 5: Comparison of lightweight mesh reconstruction methods across five scenes. Instance-level Gaussian representations optimized by SemLite-GS provide a common geometric basis. PolyFit, KSR, and Compod receive identical Gaussian-center point clouds, whereas SemLite-GS generates meshes directly from the corresponding optimized Gaussian representations. Superscripts report successful reconstructions relative to the number of target instances. Metrics are averaged over successful instances within each scene, and overall values pool successful instances across scenes. The evaluated building subsets therefore differ by method.
Figure 9: Cross-view comparison of instance masks for three representative buildings from the building and tec_school scenes. Columns correspond to six oblique views. Within each pair of rows, the upper row shows independent per-view SAM3 masks in red, and the lower row shows the masks propagated by SemLite-GS in cyan. “No target mask” indicates that no independent SAM3 prediction was associated with the target building in the corresponding view. Target association is determined using fixed projected instance labels, while the displayed mask boundaries are obtained directly from the independent SAM3 predictions.
Scene
Images
Independent SAM3 (s) ↓
SemLite-GS (s) ↓
building
3,646
1,031.3
298.0
tec_school
2,394
677.5
148.1
Table 6: Runtime of multi-view instance-mask generation on an NVIDIA RTX 4090 GPU. Independent SAM3 processes all oblique images sequentially, whereas SemLite-GS segments overlapping DOM tiles and propagates the resulting instance labels to the registered views through parallel visibility-aware rendering.
Figure 10: Visual effect of the frozen-context protection loss Lguard . The upper four rows show renderings of six representative buildings from the building and tec_school scenes, each observed from two viewpoints. Within each scene, results without the loss precede those with the loss. Red boxes and enlarged insets identify boundary leakage and spurious connections. The bottom two rows show lateral views of the corresponding local Gaussian distributions without and with Lguard , respectively. Green denotes building-associated Gaussians, and red marks spurious Gaussians near building boundaries.
CD (m) ↓
Scene
Lguard
TSDF mesh
Lightweight mesh
building (24)
Without
0.556
0.727
With
0.481
0.702
tec_school (11)
Without
0.460
0.652
With
0.445
0.616
Overall (35)
Without
0.526
0.703
Table 7: Effect of the frozen-context protection loss Lguard on the geometric accuracy of dense TSDF and lightweight building meshes. CD values are averaged over the building instances in each scene; the overall row reports the mean across all 35 instances.
Scene
Pairs ( B )
Time (s) ↓
Peak memory (GiB) ↓
building
1
6693.5
8.95
2
4867.8
10.39
4
3766.2
15.58
8
2998.0
22.86
tec_school
1
2214.2
5.25
2
1593.2
6.22
Table 8: Cumulative training-loop time and peak GPU memory for different numbers of camera–instance pairs processed in parallel. The same building instances are used at every setting; times are summed across all training groups.
Scene
Configuration
Time (s) ↓
GPU (GiB) ↓
building (24)
2×4 (CPU)
2555.2
19.41
3×4 (CPU)
2404.1
22.91
tec_school (8)
2×4 (GPU)
690.7
20.05
Table 9: Training time and peak GPU memory under multi-process pair submission. Configuration gives the number of processes × pairs per process; parentheses indicate where the inputs reside.