PlenoCI: Plenoptic CharacterIstics for View Dependence Aware Change Classification

Jason Lai; Chamuditha Jayanga Galappaththige; Niko Suenderhauf; Dimity Miller; Donald G. Dansereau

cs.CVSep 24, 2026

PlenoCI: Plenoptic CharacterIstics for View Dependence Aware Change Classification

Authors: Jason Lai, Chamuditha Jayanga Galappaththige, Niko Suenderhauf, Dimity Miller, Donald G. Dansereau

Organizations: Australian Centre for Robotics, School of Aerospace, Mechanical and Mechatronic Engineering, The University of Sydney · QUT Centre for Robotics · ARIAM Hub

Abstract

Radiance field representations such as 3D Gaussian Splatting (3DGS) natively encode complex visual phenomena such as occlusions and view dependence, but they are inherently underconstrained. Independently optimized reconstructions converge to different primitive configurations, even in unchanged regions. We introduce Plenoptic CharacterIstics (PlenoCI), a novel feature built from the plenoptic field these representations approximate. PlenoCI directly captures rich visual behaviors while ignoring Lambertian textures. By deriving closed-form analytic plenoptic derivatives from a 3DGS representation, we efficiently detect these 5D structures. Our approach is robust to underconstrained representations by construction, reporting two orders of magnitude fewer false positives between independent reconstructions of unchanged scenes than concurrent work. We demonstrate PlenoCI's utility on change classification. First, we detect changes with an instance-aware 3DGS pipeline, achieving state-of-the-art results on CL-Splats with a 25.7% mIoU gain over the strongest competitor, while remaining competitive on the more challenging PASLCD benchmark. Leveraging PlenoCI, we classify changes as geometric or appearance-based with a balanced accuracy of 0.735, comparable to the best performing baseline. We believe plenoptic derivatives and PlenoCI open new directions for view dependence aware understanding in visually complex environments. Code and data are available at https://js0n-lai.github.io/plenoci.

Figures & tables

Figure 1 : Given two 3DGS reconstructions from different times, our method first detects scene changes. To classify changes, we draw on view-dependent plenoptic structures encoded in our novel PlenoCI feature. For example, we correctly detect the sheet of paper and book as changes, and classify them using PlenoCI as occlusion edge evidence (left). Our approach is robust to the inherently underconstrained nature of 3DGS representations; two independent optimizations given the same training images yield different primitive configurations. Primitive-space comparisons [ 22 ] report false positives in static scenes while our method does not (right).

Panel of Figure 2 — Figure 2 : Example scenario in flatland to motivate our novel PlenoCI feature. Given a 2D world ( Fig. 2(a) ), we build the 3D plenoptic field $\mathscr{L}$ and compute partial derivatives $\mathscr{L}_{*}$ ( Figs. 2(b) to 2(c) ). These are useful signals for detecting visually complex phenomena such as occlusion edges ( Fig. 2(d) ). Brightness represents magnitude corresponding to each color channel’s plenoptic field. While we show ground truth geometry and $\mathscr{L}$ here, this also generalizes to radiance field representations.

Figure 3 : Our PlenoCI feature extraction pipeline. Given a 3DGS representation $\mathcal{G}$ , we sample analytic plenoptic derivatives $\nabla\mathcal{L}$ over a local light field to compute the structure tensor. The third eigenvalue $\lambda_{2}$ provides an informative signal for complex visual behaviors where existing features are unreliable, while ignoring Lambertian textures. We mask pixelwise $\lambda_{2}$ with an edge image (dilated for clarity) to yield PlenoCI $\mathcal{R}^{k}$ from view $k$ . Repeating this over selected views yields $\mathcal{R}.$

Figure 4 : Our overall scene change detection and classification approach. Given instance-aware 3DGS scene representations $\mathcal{G}_{0}$ and $\mathcal{G}_{1}$ , we aggregate per-ray change scores $\delta$ in the corresponding primitives. This yields a sparse set of anchors $\Delta\mathcal{G}^{\prime}_{t}$ to propagate on changed instances. For change detection, we sample edge rays $\mathcal{R}^{e}$ to find changed Gaussians $\Delta\mathcal{G}^{\text{all}}_{t}$ . For change classification, we leverage PlenoCI $\mathcal{R}$ to find geometrically changed Gaussians $\Delta\mathcal{G}^{\text{geo}}_{t}$ . These explicit representations can be used to render novel change masks with disambiguation between geometric and appearance changes.

Method	CL-Splats [ 1 ]		PASLCD [ 20 ]
Method	mIoU $\uparrow$	F1 $\uparrow$	mIoU $\uparrow$	F1 $\uparrow$
CYWS [ 60 ]	0.495	0.640	0.273	0.398
GeSCF [ 35 ]	0.692	0.793	0.477	0.611
SceneDiff [ 75 ]	0.334	0.433	0.473	–
3DGS-CD [ 47 ]	0.700	0.792	0.209	0.339
MV3DCD [ 20 ]	0.634	0.752	0.478	0.628

Table 1 : Binary SCD results averaged over CL-Splats [ 1 ] and PASLCD [ 20 ] . PASLCD baselines are sourced from [ 23 , 75 ] . Our method achieves state-of-the-art performance on CL-Splats, and is competitive with the strongest PASLCD baselines. The first , second , and third best performances are highlighted.

Figure 5 : Qualitative comparison of binary change masks for CL-Splats [ 1 ] (rows 1–2) and PASLCD [ 20 ] (rows 3–4). Pairwise approaches such as GeSCF [ 35 ] are susceptible to spurious detections. Our instance-aware 3DGS approach outperforms all baselines on CL-Splats, and is competitive to state-of-the-art on the challenging PASLCD benchmark.

Method	mIoU $\uparrow$	F1 $\uparrow$
No RGB	0.530	0.671
No Semantics	0.469	0.614
No Occupancy	0.514	0.658
Ours	0.538	0.688

Table 2 : Change scoring ablation averaged over PASLCD [ 20 ] . Incorporating all elements provides the best performance.

Method	mIoU $\uparrow$	F1 $\uparrow$	Precision $\uparrow$	Recall $\uparrow$
Ours	0.538	0.688	0.712	0.691
Oracle	0.596	0.731	0.667	0.839

Table 3 : SCD results averaged over PASLCD [ 20 ] varying instance segmentations. Oracle derives instances from ground truth masks while we use learned instances [ 43 ] . Recall is the main metric, improving by 20% with oracle instances as the learned instances undersegment small changed objects.

Figure 6 : Qualitative change classification results for PASLCD [ 20 ] . Top to bottom: Porch, Lounge, Meeting Room, Zen. Our PlenoCI feature enables building explicit geometry and appearance-based change representations for rendering multi-class change masks.

Metric	GS-Diff [ 22 ]	Ours	Oracle
Balanced Accuracy $\uparrow$	0.792	0.735	0.851
Geo Precision $\uparrow$	0.970	0.985	0.970
Geo Recall $\uparrow$	0.896	0.745	0.871
App Precision $\uparrow$	0.478	0.431	0.463
App Recall $\uparrow$	0.687	0.725	0.830

Table 4 : CC performance averaged over PASLCD [ 20 ] . The best performance excluding Oracle is bolded . Oracle uses instances derived from ground truth masks ( Sec. 5.2 ), driving improvements in recall of both classes. Ground truth class imbalance favoring Geo amplifies App classification errors.

Method	Binary (%) $\downarrow$	Geo (%) $\downarrow$	App (%) $\downarrow$
GS-Diff [ 22 ]	0.389	0.064	0.325
Ours	0.004	0.001	0.003

Table 5 : False positive rate from comparing two independently optimized 3DGS models of an unchanged scene, averaged over PASLCD [ 20 ] . Scenes with real changes contain 3.51% change pixels on average, ranging from 0.17–20.12%.

Figure 7 : Pixelwise eigenvalue signals derived from the plenoptic structure tensor. Given an RGB view (column 1), we sample analytic plenoptic derivatives over a local light field to build pixelwise structure tensors $M$ . The eigenvalues $\lambda_{i}$ of $M$ provide informative signals of scene structures, such as edges (column 2), corners (column 3) and occlusion edges as well as specular highlights (column 4).

Figure 8 : Propagation of change information from anchors $\Delta\mathcal{G}^{\prime}_{t}$ (column 1, highlighted in green) to produce dense change representations $\Delta\mathcal{G}_{t}$ (column 3, highlighted in green) depends on the quality of the instance-aware 3DGS (column 2). While our anchors correctly identify changed objects, upstream undersegmentation (rows 1–2) can cause our instance gating to attenuate these anchors, creating false negative predictions (column 4). With oracle segmentations (rows 3–4), our predicted changes are considerably more accurate.

Metric	All Rays	Edge Rays (Ours)
mIoU $\uparrow$	0.517	0.538
F1 $\uparrow$	0.661	0.688
Ext TFLOPs $\downarrow$	35.0	34.9
Ext VRAM (GB) $\downarrow$	10.3	5.13
Sco GFLOPs $\downarrow$	1.11	0.043
Sco VRAM (GB) $\downarrow$	3.64	1.92

Table 6 : Comparison of ray selection approaches averaged over all scenes in PASLCD [ 20 ] using 20 sample views each. Floating point operations reported for Extraction (Ext) do not include rasterization, which is shared across both methods. Retaining only edge rays (Ours) marginally improves SCD performance, but substantially reduces compute costs for change scoring (Sco).

View Overlap
Condition	Ours (%) $\downarrow$	GS-Diff (%) $\downarrow$
100% (full vs. full)	0.000	0.487
50% (half vs. half)	0.003	0.397
0% (half vs. half)	0.020	0.446
Coverage
Full vs. full	0.000	0.487

Table 7 : False positive rate from comparing two independently optimized 3DGS reconstructions of the Cantina scene from PASLCD [ 20 ] under no changes. Under varying training view overlaps and levels of coverage, our method consistently reports fewer false positives than GS-Diff [ 22 ] .

Figure 9 : Comparison of independently optimized 3DGS models of the Cantina scene from PASLCD without changes. At 100% view overlap, both models receive all training views. GS-Diff [ 22 ] reports a false positive rate of 0.487% as it directly compares primitives (inset), while we correctly report no changes. At 25% view coverage, Model B receives 5 out of 20 training views. Poor reconstruction quality causes our approach to detect false positives, albeit at a lower rate than GS-Diff.

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
GS-Pool: Object-Level Change Detection in 3D Gaussian Splatting
Oct 5, 2026Boaz Keren-Gil, James Gain, Patrick Marais
University of Cape Town
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

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

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