Online scene change detection

Learning from Automatic 3D Gaussian Splatting Edits for Online Bayesian Scene Change Detection

Learn what change looks like from edits. Update the evidence as new views arrive.

Changrui Zhu & Simon JulierUniversity College London

Final probabilities after refinement, projected onto the reference 3D scene.4 scenes · 25 viewpoints each

Why learn from edits?

Can a scene provide its own change supervision?

A shared detector needs examples from many scenes—but how much data is enough to cover the changes it might encounter? In many real-world change detection tasks, reference data is collected before the observations we want to check. That time gap lets us prepare for the specific scene in advance.

We automatically edit its 3D Gaussian reconstruction to create labelled change examples and learn from them offline. The models are ready when new views arrive, using the same procedure for every scene.

Reference 3DGS+New observation I_t→Change score & mask

Render the reference scene at the current camera pose to compare the same viewpoint. The heatmap shows the final score; thresholding at 0.5 gives the binary mask.

Porch

PASLCD / Instance 1 · View 10

Reference rendered at the same query poseFinal mask on current side

ReferenceCurrent + mask

Opacity 50%Drag ↔ to compare before & after

Outlined cyan regions show the final binary mask.Selected qualitative examples · Full method

How it works: learn from reference edits offline, then accumulate evidence as new views arrive.

Drag to compare the reference and current observation. Turn on Mask or Probability to inspect the recorded model output.

01 / Offline preparation

Learning from edits.

Edit the reference 3D Gaussian scene to generate labelled change pairs. Use these pairs and unchanged reference views to fit the scene-specific components before deployment.

Offline workflow: automatic recolouring, removal and translation with relighting, followed by fitting the scene-specific change evidence components. — Automatic reference editing and change evidence learning.Click to enlarge ↗

02 / Synthetic supervision

What does an edit look like?

Rendered edits are transferred onto real reference images to form synthetic observations. Lighting variations preserve the change labels.

I_t^synSynthetic query=I_t^refReal reference+Î_t^editEdited render−Î_tOriginal render

Add the rendered edit difference to the real image, retaining its image–render discrepancies. Pixel values are clipped to [0, 1].

Garden: real reference image before recolouring — Reference image

Garden: synthetic query with recoloured chair cushion — Synthetic query

Garden: edit mask marking the changed chair cushion in white — Training label

Recolouring changes the chair cushion. White marks the edited region; this is a training label, not a model prediction.

03 / Bayesian evidence updates

One view is not the whole story.

A strong response in one image may come from a real change, lighting, or reconstruction error. We compare each Gaussian’s response with its expected behaviour in an unchanged scene, accounting for response uncertainty. Local Bayes factors then add evidence for or against change to a shared 3D history, without replaying past images or optimizing online.

C_t: Current-view evidence
H_t: Historical support from earlier views
Q_t^full: Final refined change score
F_t: Updated 3D log-evidence field

Predict with F_t−1, then update to F_t for subsequent frames. Fitted models stay fixed online.Click to enlarge ↗

04 / Results · PASLCD

Measured across 20 scene-instances.

Our full online method improves mIoU by 18.40 percentage points over O-SCD(O) under the same local evaluation. Pixel-only results show performance before object-guided refinement.

67.64%mIoU

78.87%F1

+18.40ppmIoU vs O-SCD(O)

PASLCD · 20 scene-instances · Scores in % · Higher is better
Method	Processing	mIoU ↑	F1 ↑
Paper-reported results
SCaR-3D	Offline	19.10	28.90
3DGS-CD	Offline	20.90	33.90
MV3DCD	Offline	46.10	61.20
GS-Diff	Offline	64.40	75.80
Our evaluation
O-SCD(O)	Online	49.24	64.60
O-SCD(R)	Offline	55.99	70.27
Ours · Pixel	Online	60.36	73.80
Ours · Full	Online	67.64	78.87

Our scores average two complete runs. mIoU measures mask overlap; F1 balances precision and recall. O-SCD variants are locally re-evaluated; the other baselines are reported in the cited papers listed in our manuscript. Online/offline describes query processing; our scene-specific fitting is offline.

75.82%mIoU

85.90%F1

+25.88ppmIoU vs O-SCD(O)

Two mask protocols: published results use post-state masks; our evaluation uses pre ∪ post masks to include both vacated and newly occupied regions. Compare methods within each group.

CCS3D · 4 synthetic scenes · Scores in % · Higher is better
Method	Processing	mIoU ↑	F1 ↑
Paper-reported · Post-state masks · For context
MV3DCD	Offline	30.40	44.10
CL-Splats	Offline	39.90	53.80
3DGS-CD	Offline	41.60	51.20
SCaR-3D	Offline	65.00	72.40
Our evaluation · Union masks (pre ∪ post)
O-SCD(O)	Online	49.94	64.61
O-SCD(R)	Offline	43.59	57.82
Ours · Pixel	Online	66.83	79.47
Ours · Full	Online	75.82	85.90

Four scenes: livingroom, desk, bookcase and bedroom, weighted equally over two runs. mIoU measures mask overlap; F1 balances precision and recall. The highlighted improvement compares our full method with O-SCD(O) under the same union-mask evaluation. Online/offline describes query processing; our scene-specific fitting is offline.

Compare the detections.

Correct detectionFalse positiveMissed change

05 / Method in motion

From evidence to refined masks.

Follow a complete 25-view Porch sequence. First, combine current evidence, historical support and the fitted prior. Then compare the pixel masks with the final refined masks for the same observations.

00:00 Evidence fusion · 00:20 RefinementRecorded outputs slowed for explanation; not a runtime demonstration.

Runtime, measured separately: the full method averages 58.23 ms/frame (17.2 FPS) on RTX 5090, including the frontend; the pixel pipeline runs at 27.9 FPS. Timing covers all 25 frames across 20 PASLCD instances after initialization, including input waits and frame-to-frame overhead.