Recent open-vocabulary segmentation models have advanced semantic perception for UAVs, but predictions from moving aerial platforms can remain temporally inconsistent across repeated observations of the same physical scene. We investigate temporal semantic stability by associating frame-wise predictions with persistent world-space locations through metric 3D fusion. We introduce a voxel-level evaluation framework that jointly characterises final semantic agreement, Semantic Belief Drift (SBD), Observation Persistence (OP), and semantic uncertainty. Experiments on UAVid-3D reveal substantial frame-wise semantic flicker and show that high aggregate world-space agreement can overstate temporal stability when locations have limited repeated-observation support. Persistence-stratified analysis shows that recurrent voxels expose greater semantic disagreement, while belief drift decreases as additional evidence accumulates. This behaviour is observed across two segmentation backbones and remains consistent under variations in voxel resolution, geometric association, and temporal sampling density. Conditions that reduce world-space recurrence can increase apparent aggregate stability, demonstrating that semantic consistency must be interpreted together with observation support. Our findings highlight observation persistence as an essential conditioning variable for evaluating long-horizon semantic reliability.
Figures & tables
Figure 1: Overview of the world-space semantic fusion process for a representative UAVid-3D scene: (a) RGB frame, (b) SegEarth-OV semantic prediction rendered with the same class palette as the fused representation, and (c) world-space semantic evidence fusion obtained by projecting and accumulating semantic observations from the sampled sequence using depth and camera pose information. The fused representation serves as the analysis substrate for studying temporal semantic stability.
Table 2: Multi-scene world-space semantic stability analysis using persistent voxel evidence. VSS measures final semantic agreement, SBD measures semantic belief drift, OP denotes observation persistence, and entropy measures final semantic uncertainty.
Figure 2: Distribution of voxel-level semantic instability across the evaluated UAVid-3D scenes. Left: empirical cumulative distribution of semantic disagreement, defined as 1−VSS , where lower values indicate stronger final semantic agreement. Right: empirical cumulative distribution of Semantic Belief Drift (SBD), where lower values indicate less temporal evolution in the accumulated voxel belief. The x-axes are shown over the informative low-to-moderate instability range for readability.
Persistence
Voxels (%)
VSS ↑
SBD ↓
Entropy ↓
OP =1
57.89
1.000
–
0.000
OP =2
20.32
0.873
0.0790
0.1602
3≤ OP ≤5
18.02
0.869
0.0454
0.2285
OP >5
3.77
0.870
0.0216
0.2530
Table 3: Persistence-stratified semantic stability for SegEarth-OV at 0.2m voxel resolution. Fractions are computed over the reconstructed voxel population.
Figure 3: Qualitative visualisation of voxel-level temporal semantic stability for scene1 : (a) stability categories, (b) Semantic Belief Drift, and (c) semantic uncertainty. The visualisation shows that drift and uncertainty are spatially localised after frame-wise semantic observations are projected into world space.
Class
Voxels
VSS ↑
SBD ↓
Entropy ↓
building
286,565
0.939
0.0626
0.0464
road
207,895
0.915
0.0913
0.0685
tree
97,988
0.929
0.0851
0.0612
low vegetation
102,703
0.961
0.0393
0.0375
moving car
10,564
0.960
0.0498
0.0403
Table 4: Class-wise semantic stability across all evaluated scenes. Rare classes with very few voxels should be interpreted cautiously.