Reconstructing environments where parts of the scene change between captured image sets poses a challenge for 3D scene reconstruction. We present ChronoFuseGS, a multi-temporal Gaussian Splatting approach that addresses this issue by taking multiple separately trained Gaussian Splatting models, each representing a distinct timestep and partially overlapping in geographic coverage, and merging them into a single combined model. By allowing Gaussians from one timestep to contribute to the reconstruction at other timesteps, our approach leverages data across all captured timesteps to refine persistent parts of the scene. The model supports incremental extension, allowing new timesteps to be added while preserving the existing merged reconstruction. It encodes, for each Gaussian primitive, at which timesteps it contributes to the reconstruction. To support visual exploration of the reconstructed scene, we present a change-aware visualization approach that highlights the parts of the scene that have changed across a user-defined time selection, while preserving the color of persistent parts. Since the persistence encoding operates at the Gaussian primitive level, changes are visualized at sub-object granularity rather than being limited to object-level changes. We evaluate our approach on a real-world outdoor dataset of a flood management area, captured over 7 months across eight recording days and covering seasonal vegetation changes, snow cover, and flooding events, which we make publicly available. Our results demonstrate that the combined model consistently outperforms individually trained single-timestep models in novel-view synthesis quality, recovers structural details absent in the individual reconstructions, and reliably highlights changes in fine details and sub-parts of objects and natural structures.
Figures & tables
Figure 1: Incremental reconstruction and refinement pipeline. A set of individually trained Gaussian Splatting models and their corresponding image sets are merged into a multi-temporal model GINIT . This is followed by cross-timestep initialization to estimate the per-Gaussian persistence encoding oi and light compensation li , training Gaussians G\textscinitk originating from Gk against Ip=k , and by a combined refinement step. The refined model enables novel-view synthesis and change highlighting H(t) for any user-defined time selection t . Solid arrows indicate the initial computation of G′ , dashed arrows the incremental fusion of an additional model Gq into G′ , yielding G′′ .
Figure 2: Individual timestep renderings and change highlighting for different time selections. (a)–(c) show the multi-temporal model rendered at t1 , t3 , and t4 ; (d) shows the blended multi-time view ( t1 , t3 , and t4 ); (e)–(f) show change highlights with Gaussian color ci converted to grayscale. In (e), the scaffolding is absent at t1 , so it is rendered in the combined highlight color of t3 and t4 . In (f), the time selection is restricted to the last two timesteps, and the scaffolding appears in grayscale as it changed minimally between them. ∙ indicates enabled timesteps, with color corresponding to timestep presence in (e)–(f); ∘ indicates a disabled timestep.
Figure 3: Reconstruction of the Baluster Vase dataset across four timesteps, with new cracks progressively added to the central vase. (a) Blended multi-time view with all timesteps selected; (b) the same view with change highlighting, where persistent parts are rendered in grayscale and crack colors indicate at which timesteps they are present. ∙ indicates an enabled timestep, with color corresponding to timestep presence in (b).
Figure 4: Reconstruction of the Coral data with change visualization and different time selections. The Gaussian color ci is converted to grayscale. In (a), the last two timesteps are enabled, and in (b), changes across all four timesteps are highlighted. ∙ indicates an enabled timestep, with color corresponding to timestep presence; ∘ indicates a disabled timestep.
Figure 5: Individual timestep renderings and change highlighting for different time selections. (a) Individual timestep renderings at t1 , t2 , and t3 . (b) t=[1,1,0] , with little change between the first two timesteps, resulting in a nearly unaltered representation. (c) t=[1,1,1] , extending the selection to all three timesteps, highlighting changes across all timesteps. ∙ indicates an enabled timestep, with color corresponding to timestep presence; ∘ indicates a disabled timestep.
Figure 6: Different gradients for change highlighting across four selected timesteps.
Name
T
Days
Images
No Wolf in the Meadow (NWM)
Complete
9
8
2221
Autumn
4
4
975
Flooding
3
3
629
Snow
3
1
608
Car
3
1
705
Table 1: Overview of datasets, including the number of timesteps T, recording days, and total input images. The snow subset includes two timesteps that are not present in the complete dataset. Some timesteps are used for multiple subsets. Synthetic datasets (Baluster Vase, Corals) have no associated recording days.
Figure 7: Reconstruction results on five subsets of the No Wolf in the Meadow dataset. Rows correspond to dataset subsets; columns show the best-performing individually trained baseline, ChronoFuseGS at 30k refinement iterations, and the ground-truth test image.
Figure 8: Results of our approach with 30,000 refinement iterations for different datasets, where (a) and (b) share the shown timestep. The Autumn subset includes 4 timesteps, and the Complete NWM dataset includes 9 timesteps.
Complete (4.67)
Autumn (2.17)
Flooding (1.65)
Snow (0.81)
Car (1.56)
Method
PS ↑
SS ↑
LP ↓
PS ↑
SS ↑
LP ↓
PS ↑
SS ↑
LP ↓
PS ↑
SS ↑
LP ↓
PS ↑
SS ↑
LP ↓
Individual Init
21.6
0.73
0.45
20.9
0.69
0.47
22.2
0.76
0.43
23.2
0.82
0.38
22.7
0.76
0.44
Ours Merged
21.6
0.73
0.45
20.9
0.69
0.47
22.2
0.76
0.43
23.2
0.82
0.38
22.7
0.76
0.44
Ours 6k
21.5
0.71
0.44
21.2
0.70
0.46
22.8
0.78
0.41
23.3
0.82
0.37
23.1
0.78
0.42
Ours 15k
21.8
0.73
0.43
21.1
0.69
0.47
22.8
0.78
0.41
23.4
0.82
0.36
23.2
0.78
0.42
Ours 30k
21.9
0.74
0.43
21.3
0.70
0.46
22.8
0.78
0.40
23.4
0.83
0.36
23.3
0.79
0.41
Table 2: Quantitative comparison across the five subsets of the No Wolf in the Meadow dataset. Numbers in parentheses indicate the combined number of Gaussians of the input Gaussian sets in millions. PS: PSNR, SS: SSIM, LP: LPIPS. ↑ indicates higher is better, ↓ lower is better. Best , second best , and third best results are highlighted per column. Individual Init represents the individual trained Gaussian sets used as input for our approach, and Ours Merged represents the multi-temporal model after merging the individual Gaussian sets before additional training.
Training Time (min)
Rendering FPS
Dataset
Merge
Cross
Ref
Sum
C
H
H0,1
NWM
Complete
75.4
48.3
12.0
135.7
48
49
49
Autumn
19.3
14.4
4.5
38.2
120
120
121
Flooding
11.3
7.5
4.0
22.7
149
151
151
Snow
15.2
6.0
3.0
24.2
231
233
232
Table 3: \added Training times in minutes for 6000 refinement iterations and average rendering FPS of our Unreal implementation. Merge, Cross, and Ref denote the pipeline stages: Merg ing to Multi-Temporal Model, Cross -Timestep Initialization, and Combined Ref inement. C and H represent the FPS with all timesteps selected in true c olor and change h ighlighting respectively. H0,1 represents the FPS for change highlighting with only the first two timesteps selected.
Figure 9: \added Training time up to each timestep of the NWM (complete) dataset, comparing incremental fusion—where the multi-temporal model is extended with a new timestep—against retraining from scratch, where the model is fully recomputed from all timesteps available so far. Each bar represents the time to incorporate the new timestep and yield a single refined model containing all timesteps from t0 through the new one. Times broken down by pipeline stage: merging, cross-timestep initialization, and refinement.
Figure 10: Screenshots of the web survey tool for the three conditions.
Figure 11: Quantitative results of the user study across three conditions (I: Interactive, S: Static, B: Baseline). The black line indicates the median.
Figure 12: Change visualization between two timesteps for different scene change types. Insets show the individual timestep renderings. (a) Rocks partially covered by vegetation at different levels. (b) Scene covered with ice at one timestep, with additional objects present during the other timestep. (c) River with differing water levels, highlighting the change in waterline. ∙ indicates an enabled timestep, with color corresponding to timestep presence; ∘ indicates a disabled timestep.
Figure 13: Demonstration of our change visualization on regions not covered by the input data at one timestep. (a) Yellow highlights areas not covered at t3 but present at t1 and t2 ; grayscale indicates regions persistent across all timesteps. (b)–(d) Individual timestep renderings for t1 – t3 . ∙ indicates an enabled timestep, with color corresponding to timestep presence in (a); ∘ indicates a disabled timestep.
Recent extensions of 3D Gaussian Splatting (3DGS) enable real-time novel view synthesis in dynamic scenes by learning time-conditioned Gaussian deformations. However, existing MLP-based methods typically estimate deformations independently at each timestamp, making them less robust to large or abrupt motions. To address this issue, we propose EvoGS, a 3DGS-based dynamic reconstruction framework that models Gaussian deformation as a temporal evolution process. EvoGS maintains persistent deformation states for each Gaussian, extrapolates future states from historical deformation states, and corrects the predictions with MLP-derived observations. The correction is adaptively weighted using a temporal residual memory and evolution statistics such as deformation velocity and trajectory deviation. To further improve reconstruction quality, EvoGS introduces deformation-aware densification. Clone and split operations are performed along corrected deformation directions, while an uncertainty-aware strategy suppresses densification for Gaussians with unstable deformation histories. Experiments show that EvoGS improves dynamic novel view synthesis quality and achieves competitive performance across benchmarks.
Wei Dong, Shahram Shirani, Jun Chen +1
The Department of Electrical & Computer Engineering, McMaster University
Factories, museums and surveyors photograph the same space months apart and need to know which objects changed. When each visit is reconstructed with 3D Gaussian Splatting (3DGS), a direct comparison of the two reconstructions does not answer this. Training is stochastic, so two reconstructions of an unchanged space never coincide, and the second visit is often a quick re-scan with far fewer photographs. We propose GS-Pool, which takes two independently reconstructed Gaussian fields of the same space and returns the changed objects in each, together with their masks. SAM2 masks of each visit's photographs are lifted onto the Gaussians that render them and merged into an object pool, so every decision is taken once per object in 3D. We introduce a photographic carrier, the 3DGS training loss of each input reconstruction against the other visit's photographs, backpropagated to the Gaussians that rendered each pixel. We combine it with GS-Diff's geometry and colour terms and our distilled DINOv3 features. This evidence is compared with that of the objects present in both visits, which sets a change threshold for each scene. On PASLCD, GS-Pool reaches mIoU/F1 scores of 0.751/0.846 against 0.644/0.758 for GS-Diff, the strongest prior method, a gain of 17%/12%. Its mIoU is also 36%, 40% and 57% above that of O-SCD, PlenoCI and MV-3DCD, and it reaches 0.855 mIoU on CL-Splats, 33% above MV-3DCD. Each changed object is returned as a set of Gaussians with the evidence behind its decision, which an inspector can review in 3D.
Dynamic 3D Gaussian Splatting (3DGS) methods reconstruct time-varying scenes from synchronized multi-camera video using photometric supervision. When a moving object becomes fully occluded from all training cameras, this supervision vanishes: the Gaussians representing it receive no gradient signal and degrade. Existing approaches to incomplete observations in neural reconstruction rely on learned generative priors that prioritize visual plausibility over physical correctness. We propose PersistGS, a method that restores object permanence during occlusion by coupling differentiable rigid body simulation with 3D Gaussian Splatting. Our approach decomposes the scene into per-object Gaussians and collision meshes, estimates friction and velocity from the observed pre-occlusion trajectory via differentiable simulation, and uses the resulting SE(3) trajectory to position object Gaussians throughout the occlusion period. Because the predicted trajectory satisfies the governing equations of rigid body dynamics, it faithfully captures contact events (bounces, friction-based deceleration, direction changes) that kinematic extrapolation cannot model. We introduce a centroid silhouette loss that isolates positional gradients from appearance noise, yielding 40% lower trajectory error than photometric supervision. We evaluate using cameras withheld from training that observe the object during its occlusion. Experiments on synthetic scenes show that PersistGS outperforms constant velocity extrapolation by +2.46dB PSNR and comes within 0.19dB of a ground-truth trajectory upper bound.