GS-Pool: Object-Level Change Detection in 3D Gaussian Splatting

Boaz Keren-Gil; James Gain; Patrick Marais

cs.CVOct 5, 2026

GS-Pool: Object-Level Change Detection in 3D Gaussian Splatting

Authors: Boaz Keren-Gil, James Gain, Patrick Marais

Organizations: University of Cape Town

Abstract

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.

Figures & tables

Panel of Figure 2 — Figure 2 : GS-Pool on one PASLCD cell: a distilled DINOv3 field per visit (1), lifted SAM2 segments merged into objects (2), four carriers per primitive and the scene’s null (3), one decision per object (4), and the change mask and changed primitives (5).

Figure 3 : A glass emptied in place on Lounge (a). Geometry and colour rank it near the scene’s middle (b, c), semantics higher (d) and the photographs near the top (e). (f) GS-Pool ’s mask and the annotation.

PASLCD
			mean		mIoU per reconstruction			F1 per reconstruction
Method	Frame	Annotations	mIoU	F1	seed 22	23	24	seed 22	23	24
GS-Pool (FastPGSR)	joint	PASLCD	0.7509	0.8461	0.7404	0.7520	0.7602	0.8379	0.8466	0.8537
GS-Pool (vanilla 3DGS)	joint	PASLCD	0.7198	0.8209	0.7223	0.7111	0.7259	0.8236	0.8144	0.8248
GS-Pool (vanilla 3DGS)	anchored	PASLCD	0.7281	0.8266	0.7368	0.7139	0.7335	0.8341	0.8159	0.8300
GS-Pool (FastPGSR)	anchored	PASLCD	0.7260	0.8277	0.7135	0.7347	0.7300	0.8171	0.8351	0.8309

Table 1: Change detection on PASLCD and CL-Splats. GS-Pool is given as the seed mean and per seed. Baselines are as reported ( MV-3DCD on PASLCD as reported in [ 22 ] ), except the rows on our CL-Splats annotations, which we ran ( † four of five scenes).

Figure 4 : Porch at the test view with the most annotated change: the annotation, GS-Pool on FastPGSR and on vanilla 3DGS, and O-SCD , each tagged with its view IoU.

Figure 5 : The output in 3D on Porch from a viewpoint no camera took. Each visit’s changed primitives are shown on its own reconstruction and the rest in grey: $T_{1}$ ’s two stools and mug (a), and $T_{2}$ ’s stool and the objects on its table (b).

		FastPGSR		vanilla 3DGS
row	decision	mIoU	F1	mIoU	F1
1	photographs alone	0.4681	0.5896	0.4353	0.5627
2	+ semantics	0.5899	0.7087	0.6060	0.7291
3	all four carriers, averaged	0.6090	0.7252	0.6107	0.7357
4	core, $F(o)>\theta$	0.7034	0.8080	0.6617	0.7762
5	+ photo rank, quorum	0.7113	0.8147	0.6668	0.7796

Table 2: The decision rule built up one component at a time on the same pools and carriers. Mean over the 60 joint-frame members per trainer. Row 6 is the full rule.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7 : SAM2’s stability threshold on the joint frame, mean over 60 members per trainer. Dashed: the chosen

0.90

. Dotted: the library’s default

0.95

row	prims/visit	time
FastPGSR, joint	0.19M	2.1 min
FastPGSR, anchored	0.17M	1.7 min
vanilla 3DGS, joint	1.14M	5.7 min
vanilla 3DGS, anchored	1.13M	4.2 min

Appendix

Table 3 : Median wall-clock per member on one RTX 4090 (24 GB), from the reconstructions, distilled fields, cached features and segmentation maps to the masks and change PLYs. Producing those inputs is excluded. Stage 3’s neighbour searches run on the GPU.

GS-Pool , re-scored by this pass	0.7509	0.8461
	mIoU	F1
oracle: greedy decisions, same pool	0.8448	0.9090
oracle: every false pixel removed	0.8172	0.8882
oracle: every missed pixel filled	0.9229	0.9568

Appendix

Table 4: The decision oracle and the two pixel bounds on the same 60 members (FastPGSR, joint frame). The first row re-scores GS-Pool with the same pass and reproduces its score exactly.

		FastPGSR		vanilla 3DGS
	comparison	mIoU	F1	mIoU	F1
	full rule	0.7509	0.8461	0.7198	0.8209
	core alone (row 4)	0.7034	0.8080	0.6617	0.7762
null	full rule, own-visit null	0.7464	0.8407	0.7269	0.8205
	full rule, both visits’ null	0.7357	0.8301	0.7195	0.8170
	core alone, own-visit null	0.6530	0.7647	0.5986	0.7109

Appendix

Table 5: Twelve controlled comparisons around the full rule, each changing one thing, over the 120 joint members.

Method	mean	median	scenes below MV-3DCD
MV-3DCD [ 16 ]	1.72	1.34	—
O-SCD [ 22 ] †	1.47	1.09	3 of 9
GS-Pool (FastPGSR)	1.03	0.31	6 of 10
GS-Pool (vanilla 3DGS)	1.15	0.17	7 of 10

Appendix

Table 6 : Pixels flagged between two captures of the same unchanged scene under different lighting (%), at instance 2’s reference cameras ( † nine of ten scenes).

Method	mean	median	runs at zero
GS-Diff [ 4 ] ∗	0.389	—	—
PlenoCI [ 3 ] ∗	0.004	—	—
GS-Pool (FastPGSR)	0.153	0.017	14 of 60
GS-Pool (vanilla 3DGS)	0.511	0.017	15 of 60

Appendix

Table 7: Pixels flagged as changed between two reconstructions of an unchanged scene (%). ∗ As reported by PlenoCI [ 3 ] on its own reconstructions, not measured on our pairs.

Method	annotations	mIoU	F1
GS-Diff [ 4 ]	PlenoCI	0.724	0.829
O-SCD [ 22 ]	PlenoCI	0.756	0.845
MV-3DCD [ 16 ]	PlenoCI	0.634	0.752
PlenoCI [ 3 ]	PlenoCI	0.951	0.975
MV-3DCD [ 16 ]	GS-Pool	0.6413	0.7667
O-SCD [ 22 ] †	GS-Pool	0.6224	0.7428

Appendix

Table 8 : Change detection on the CL-Splats real scenes, PlenoCI’s figures on its own annotations [ 3 ] above and ours below ( † four scenes, 25 views).

Panel of Figure 10 — Figure 10 : All five CL-Splats scenes at the view with the most annotated change, both trainers at seed 22, labelled with the view’s IoU and the scene’s three-seed mIoU (FastPGSR / vanilla 3DGS).

Figure 12 : Cells 1–10 (Cantina to Meeting_room) at the test view with the most annotated change, joint frame, seed 22, labelled with the view’s IoU and the cell’s three-seed mIoU (FastPGSR / vanilla 3DGS).

Figure 13 : Cells 11–20 (Playground to Zen), as in Figure 12 .

Explore similar work

From Pixels to Primitives: Scene Change Detection in 3D Gaussian Splatting
May 8, 2026Chamuditha Jayanga Galappaththige, Jason Lai, Timothy Patten +3Change Detection Gaussian Splatting
1QUT Centre for Robotics · 2ARIAM · 3ACFR, University of Sydney +1
PersistGS: Differentiable Physics for Object Permanence in 4D Gaussian Splatting
Jun 2, 2026Adrian Ramlal, John S. ZelekDynamic 3D Gaussian Splatting Gaussian Splatting
University of Waterloo
OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives
May 19, 2026Guiyu Liu, Niklas Vaara, Janne Mustaniemi +23D Gaussian Opacity
Center for Machine Vision and Signal Analysis, University of Oulu, Finland · Aalto University, Finland

May 8, 2026, cs.CVJ/K move · Enter open · S save

From Pixels to Primitives: Scene Change Detection in 3D Gaussian Splatting

Chamuditha Jayanga Galappaththige, Jason Lai, Timothy Patten, Donald Dansereau+2

1QUT Centre for Robotics · 2ARIAM · 3ACFR, University of Sydney+1

Change Detection Gaussian Splatting Scene Reconstruction Pixels Partial Observability

Open paper AskPDF arXiv

Scene change detection methods built on Gaussian splatting universally follow a render-then-compare paradigm: the pre-change scene is rendered into 2D and compared against post-change images via pixel or feature residuals. This change detection problem with Gaussian Splatting has been treated as a question about pixels; we treat it as a question about primitives. We provide direct evidence that native primitive attributes alone -- position, anisotropic covariance, and color -- carry sufficient signal for scene change detection. What makes primitive-space comparison hard is the under-constrained nature of Gaussian splatting representation: independent optimizations yield primitive solutions whose count, positions, shapes, and colors differ even where nothing has changed. We address this challenge with anisotropic models of geometric and photometric drift, complemented by a per-primitive observability term that reflects the extent to which each Gaussian is constrained by the camera geometry. Operating directly on primitives gives our method, GD-DIFF, two properties that distinguish it from render-then-compare methods. First, change maps are multi-view consistent by construction, where prior work had to learn this through an additional optimization objective. Second, geometric and appearance changes are scored separately, identifying not just where but what kind of change occurred, distinguishing structural changes (e.g., an added object) from surface-level ones (e.g., a color change) without supervision or external model dependencies. On real-world benchmarks, GS-DIFF surpasses the prior state-of-the-art approach by

\sim

17% in mean Intersection over Union.

Abstract

Explore similar work

From Pixels to Primitives: Scene Change Detection in 3D Gaussian Splatting

PersistGS: Differentiable Physics for Object Permanence in 4D Gaussian Splatting

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives