TL;DRO-SCD++ folds each new view into the map in a single pass, with no replay of past frames and no iterative optimization. It is 18% more accurate (mIoU) and 2.2× faster than O-SCD (over 25 FPS), and it sets its own decision threshold without labels, with a per-pixel confidence.
Fly-throughs of O-SCD++ results on PASLCD. Top: the pre-change reference map and the post-change scene, rendered from the updated map along a smooth path through the inference poses. Bottom: the O-SCD++ change prediction (highlighted) and its decision margin (bright = low confidence).
O-SCD (iterations per frame)O-SCD++ (replay-free)O-SCD offline
Figure 1. O-SCD fuses multi-view evidence by iterative optimization over replayed cue maps, which trades detection quality against speed (blue, labelled with the per-frame iteration budget). The replay-free formulation of O-SCD++ (red) removes this trade-off: it is simultaneously more accurate and more than twice as fast, and surpasses even O-SCD's offline result while remaining strictly causal.
0.575mIoU
Online on PASLCD, up from 0.489 and above O-SCD offline (0.556)
25.6FPS
2.2× the throughput of O-SCD (11.8 FPS)
2.1ms
Multi-view fusion per frame, down from 47.8 ms
0labels
Exact label-free threshold, plus a per-pixel decision margin
Abstract
Removing the accuracy–speed trade-off
Online scene change detection requires an agent to detect what has changed in a previously mapped environment while it is still on site, observing from unconstrained viewpoints and without access to future observations. This causal setting is strictly harder than its offline counterpart, and online methods have consistently trailed offline accuracy. O-SCD closed that gap by treating change as a persistent property of a 3D Gaussian splatting map of the reference scene, surpassing the strongest offline methods in change detection accuracy while operating at over 10 frames per second. O-SCD optimizes a change field over the Gaussians of the reference map, refitting it to all past cue maps at every frame, a step that consumes more than half of the per-frame budget and forces accuracy to be traded against speed.
We present a replay-free formulation that eliminates this trade-off. Because the reference map stays fixed, rendering the change field is a linear operation. Each new observation can therefore be carried onto the map's Gaussians directly, by a single backward pass of the renderer, which reduces multi-view fusion to an exact two-scalar recursion per Gaussian. Past observations are never stored or revisited. The self-supervised objective that O-SCD minimizes by descent also determines the decision threshold exactly, by sorting, yielding a label-free per-frame cutoff whose margin provides a selective-prediction signal. On PASLCD, our method improves mean intersection over union from 0.489 to 0.575 and throughput from 11.8 to 25.6 frames per second, even exceeding O-SCD's offline result while remaining strictly causal.
Method
Every iterative step becomes a direct computation
O-SCD++ keeps the front end of O-SCD (pose estimation against the reference map and change-cue extraction) and replaces its change stage.
Click to open full resolution · equation, section and proposition numbers refer to the paperFigure 2. Overview of O-SCD++ for one inference frame. The front end, adopted from O-SCD, registers the inference image against the reference map and computes a change cue. The replay-free change stage transfers the cue onto the reference primitives with one rasterizer backward pass, folds it into a two-scalar state per primitive by an exact recursion, predicts a score for the current view from the updated state and the current cue, and binarizes it with the exact minimizer of a label-free objective, which also yields a per-pixel decision margin. It eliminates the replay-based optimization of O-SCD (dashed).
O-SCD (iterative)
O-SCD++ (direct)
(i)Transfer
Gradient steps on the self-supervised loss, backpropagated through the rasterizer
One rasterizer backward pass applies the rendering adjoint A⊤
(ii)Combine
Replay: n = 16 steps per frame on stored cue maps c1, …, ct
Exact recursion on two scalars (qg, Wg) per Gaussian
(iii)Decide
Fixed cutoff 0.5; no decision confidence
Exact minimizer of the loss by sort and scan, with a per-pixel margin μt
Costper frame
n(Trast + Tadj) + Trast memory O(G + tN)
2Trast + Tadj + O(N log N) memory O(G)
PASLCDonline
0.489 mIoU · 11.8 FPS
0.575 mIoU · 25.6 FPS
The per-frame cost loses its iteration budget n and its dependence on the number t of frames processed. G: primitives; N: pixels.
ExactnessOrder-invariantThe state reached after a sequence does not depend on the order in which the frames arrive.
ExactnessRetrospective re-inferenceRe-applied to every frame, the final state improves on the online masks without any refinement.
ExactnessDecision marginsThe exact decision yields a per-pixel margin: to our knowledge the first confidence signal from a label-free online SCD method.
Qualitative results
O-SCD vs. O-SCD++
Drag the slider to compare the change masks of O-SCD (left) and O-SCD++ (right), drawn over the post-change image.
O-SCDO-SCD++IoU 0.486 · F1 0.654IoU 0.815 · F1 0.898
Pre-change (rendered)
Post-change
O-SCDO-SCD++IoU 0.440 · F1 0.611IoU 0.764 · F1 0.866
Pre-change (rendered)
Post-change
O-SCDO-SCD++IoU 0.472 · F1 0.641IoU 0.739 · F1 0.850
Pre-change (rendered)
Post-change
O-SCDO-SCD++IoU 0.478 · F1 0.647IoU 0.744 · F1 0.853
Pre-change (rendered)
Post-change
O-SCDO-SCD++IoU 0.532 · F1 0.695IoU 0.710 · F1 0.831
Pre-change (rendered)
Post-change
O-SCDO-SCD++IoU 0.485 · F1 0.653IoU 0.662 · F1 0.796
Figure 3. Qualitative comparison with O-SCD on PASLCD, colour-coded against the ground truth, with per-frame IoU and F1. O-SCD++ also recovers particularly challenging surface-level changes that barely contrast with their surroundings: a grey cross on a grey chair (Garden), a white laminated sheet on a white table (Meeting room), and a water spill and a transparent cup on a table (Lunch room).
Quantitative results
State of the art in accuracy and throughput
LF: label-free, PA: pose-agnostic, MV: multi-view. Best in bold, second underlined; rows in grey italics are shown for reference.
PASLCD · online Averaged over all 20 instances
Method
LF
PA
MV
mIoU ↑
F1 ↑
FPS ↑
O-SCD (offline)
✓
✓
✓
0.556
0.700
–
ChangeSim (CS)
–
–
–
0.018
0.034
11.5
CS+CYWS2D
–
–
–
0.243
0.360
8.2
CS+GeSCF
✓
–
–
0.181
0.270
<1
FastGeSCF
–
–
–
0.122
0.198
1.43
SplatPose
✓
✓
–
0.173
0.281
<1
SplatPose+
✓
✓
–
0.237
0.358
<1
O-SCD
✓
✓
✓
0.489
0.643
11.8
O-SCD++
✓
✓
✓
0.575
0.715
25.6
PASLCD · offline Offline methods use the complete post-change capture. O-SCD++ re-infers every frame from its final state (retrospective inference), with no further optimization.
Method
LF
Setting
mIoU ↑
F1 ↑
CYWS2D
–
offline
0.273
0.398
VSCD
–
offline
0.112
0.186
GeSCF
✓
offline
0.477
0.611
SplatPose+
✓
offline
0.237
0.358
SCAR-3D
✓
offline
0.191
0.289
3DGS-CD
✓
offline
0.209
0.339
MV3DCD
✓
offline
0.478
0.628
SceneDiff
✓
offline
0.473
0.609
O-SCD
✓
offline
0.556
0.700
O-SCD++
✓
retrospective
0.586
0.726
O-SCD++
✓
online
0.575
0.715
CL-Splats and ChangeSim Averaged over scenes
Method
CL-Splats
ChangeSim
mIoU ↑
F1 ↑
mIoU ↑
F1 ↑
ChangeSim (CS)
0.010
0.019
0.130
0.212
CS+CYWS2D
0.372
0.519
0.271
0.408
CS+GeSCF
0.520
0.630
0.348
0.484
FastGeSCF
0.183
0.252
0.315
0.458
SplatPose
0.224
0.320
0.061
0.108
SplatPose+
0.556
0.700
0.115
0.186
O-SCD
0.657
0.773
0.404
0.535
O-SCD++
0.703
0.807
0.474
0.605
Per-frame runtime PASLCD, RTX 4090, about 1008×567 (scene-dependent)
Module
O-SCD
O-SCD++
ms
%
ms
%
Pose estimation
30.4
36.0
30.2
77
Change detection
54.4
64.0
9.0
23
Change cue generation
6.1
7.2
5.9
15.0
Multi-view fusion
47.8
56.5
2.1
5.5
Mask inference
0.5
0.3
1.0
2.5
Even online, O-SCD++ exceeds every offline method. Replay-free fusion is about 20× faster than O-SCD's, so change detection is no longer the bottleneck. All numbers on this page are from the O-SCD++ paper.
Selective prediction
Decision margins: knowing where it might be wrong
The margin is small (bright) where a slightly different threshold would flip the label. These regions coincide with misclassified pixels, change boundaries and missed regions, so a system can report only confident changes or revisit uncertain ones while still on site. Drag the slider to compare the prediction (left) with its margin map (right).
Pre-change (rendered)
Post-change
O-SCD++ predictionDecision marginIoU 0.624 · F1 0.769
Pre-change (rendered)
Post-change
O-SCD++ predictionDecision marginIoU 0.778 · F1 0.875
Pre-change (rendered)
Post-change
O-SCD++ predictionDecision marginIoU 0.610 · F1 0.758
Pre-change (rendered)
Post-change
O-SCD++ predictionDecision marginIoU 0.687 · F1 0.815
Figure 4. Decision margins of O-SCD++ on PASLCD. The margin map is bright where the margin is small, that is, where a slightly different threshold would change the label. The prediction is colour-coded against the ground truth: white true positive, black true negative, red false negative, purple false positive.
@article{galappaththige2026oscdpp,
title = {O-SCD++: Replay-Free Online Scene Change Detection via Recursive Evidence Fusion},
author = {Galappaththige, Chamuditha Jayanga and Lai, Jason and Windrim, Lloyd and Patten, Timothy and Dansereau, Donald G. and Suenderhauf, Niko and Miller, Dimity},
journal = {arXiv preprint arXiv:2610.24462},
year = {2026}
}
If you use PASLCD, please also cite MV3DCD; if you build on the online setting, please also cite O-SCD.
This work was supported by the Australian Research Council Research Hub in Intelligent Robotic Systems for Real-Time Asset Management (ARIAM, IH210100030) and Abyss Solutions. C.J.G., N.S. and D.M. also acknowledge ongoing support from the QUT Centre for Robotics.
O-SCD 
Click to open full resolution · equation, section and proposition numbers refer to the paper
O-SCD
O-SCD++
IoU 0.486 · F1 0.654
IoU 0.815 · F1 0.898
O-SCD
O-SCD++
IoU 0.440 · F1 0.611
IoU 0.764 · F1 0.866
O-SCD
O-SCD++
IoU 0.472 · F1 0.641
IoU 0.739 · F1 0.850
O-SCD
O-SCD++
IoU 0.478 · F1 0.647
IoU 0.744 · F1 0.853
O-SCD
O-SCD++
IoU 0.532 · F1 0.695
IoU 0.710 · F1 0.831
O-SCD
O-SCD++
IoU 0.485 · F1 0.653
IoU 0.662 · F1 0.796
O-SCD++ prediction
Decision margin
IoU 0.624 · F1 0.769
O-SCD++ prediction
Decision margin
IoU 0.778 · F1 0.875
O-SCD++ prediction
Decision margin
IoU 0.610 · F1 0.758
O-SCD++ prediction
Decision margin
IoU 0.687 · F1 0.815
