O-SCD++: Replay-Free Online Scene Change Detection via Recursive Evidence Fusion
Authors: Chamuditha Jayanga Galappaththige, Jason Lai, Lloyd Windrim, Timothy Patten, Donald G. Dansereau, Niko Suenderhauf, Dimity Miller
Organizations: QUT Centre for Robotics, Queensland University of Technology, Australia · Australian Research Council Research Hub in Intelligent Robotic Systems for Real-Time Asset Management (ARIAM) · Australian Centre for Robotics, The University of Sydney, Australia · Abyss Solutions
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.
Figures & tables
Fig. 1: The state-of-the-art O-SCD [ 10 ] fuses multi-view evidence by iterative optimization over replayed cue maps, which forces detection quality to be traded against speed (blue, annotated with the per-frame iteration budget). Our novel replay-free formulation (red star) derived in Sec. IV eliminates this trade-off: it is simultaneously more accurate and more than twice as fast. Our formulation surpasses even O-SCD’s offline performance while operating strictly causally.
Fig. 2: The three steps of the change stage in O-SCD [ 10 ] and in O-SCD++. Each iterative step of O-SCD is replaced by a direct computation; the per-frame cost loses its iteration budget n and its dependence on the number t of frames processed.
Fig. 3: Overview of O-SCD++ for one inference frame. The front end, adopted from O-SCD [ 10 ] , 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).
Fig. 4: Qualitative comparison with O-SCD [ 10 ] on PASLCD [ 5 ] . Each row shows the reference map rendered at the estimated pose of the inference image, the post-change inference image, and the change masks of O-SCD and O-SCD++, 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, transparent cup on a table (Lunch room).
Method
LF
PA
MV
mIoU
F1
FPS
O-SCD (offline) [ 10 ]
✓
✓
✓
0.556
0.700
–
ChangeSim (CS) [ 11 ]
–
–
–
0.018
0.034
11.5
CS+CYWS2D [ 4 ]
–
–
–
0.243
0.360
8.2
CS+GeSCF [ 23 ]
✓
–
–
0.181
0.270
< 1
FastGeSCF [ 43 ]
–
–
–
0.122
0.198
1.43
SplatPose [ 32 ]
✓
✓
–
0.173
0.281
< 1
TABLE I: SCD results on PASLCD [ 5 ] , averaged over all 20 instances. LF: label-free; PA: pose-agnostic; MV: multi-view. The offline O-SCD result is included for reference. O-SCD++ achieves the state of the art perform in both accuracy and throughput.
CL-Splats
ChangeSim
Method
mIoU
F1
mIoU
F1
ChangeSim (CS) [ 11 ]
0.010
0.019
0.130
0.212
CS+CYWS2D [ 4 ]
0.372
0.519
0.271
0.408
CS+GeSCF [ 23 ]
0.520
0.630
0.348
0.484
FastGeSCF [ 43 ]
0.183
0.252
0.315
0.458
SplatPose [ 32 ]
0.224
0.320
0.061
0.108
TABLE II: SCD results on CL-Splats [ 52 ] and ChangeSim [ 11 ] , averaged over scenes. O-SCD++ achieves the best change detection accuracy on both.
Variant
mIoU
F1
Current-view prediction ( Sec. IV-E )
Cue only (no fusion)
0.454
0.595
Map only
0.526
0.676
Map and cue (ours)
0.575
0.715
Aggregation weight ( Sec. IV-D )
Unbounded mass, ρg,t=mg,t
0.562
0.701
TABLE III: Ablations of O-SCD++ on PASLCD [ 5 ] . Each block varies one component while the others are fixed. The oracle selects, per frame, the threshold that maximizes IoU against the ground truth and serves as an upper bound.
Indoor
Outdoor
Method
mIoU
F1
mIoU
F1
O-SCD (offline) [ 10 ]
0.555
0.706
0.557
0.694
CS+CYWS2D [ 4 ]
0.167
0.247
0.319
0.448
SplatPose+ [ 45 ]
0.213
0.326
0.260
0.390
O-SCD [ 10 ]
0.484
0.641
0.494
0.645
O-SCD++
0.573
0.718
0.576
0.713
TABLE IV: Indoor and outdoor scenes of PASLCD [ 5 ] for the two strongest online baselines of Table I . O-SCD++ along with O-SCD perform similarly in both indoor and outdoor scenes.
Similar lighting
Different lighting
Method
mIoU
F1
mIoU
F1
O-SCD (offline) [ 10 ]
0.580
0.718
0.532
0.681
CS+CYWS2D [ 4 ]
0.258
0.378
0.229
0.342
SplatPose+ [ 45 ]
0.236
0.357
0.237
0.359
O-SCD [ 10 ]
0.506
0.656
0.472
0.630
O-SCD++
0.602
0.736
0.548
0.694
TABLE V: Results under lighting similar to, and different from, the reference capture, for the two strongest online baselines of Table I on PASLCD [ 5 ] . O-SCD++ achieves the best accuracy under both lighting conditions.
Method
Setting
Extra optimization
mIoU
F1
O-SCD [ 10 ]
online
–
0.489
0.643
O-SCD [ 10 ]
retrospective
none
0.511
0.662
O-SCD [ 10 ]
offline
refinement
0.556
0.700
O-SCD++
online
–
0.575
0.715
O-SCD++
retrospective
none
0.586
0.726
TABLE VI: Online, retrospective and offline inference on PASLCD [ 5 ] . Retrospective inference re-infers every frame from the state reached after the last frame; offline O-SCD instead continues to refine its change representation up to 3000 iterations.
Fig. 5: Retrospective inference with O-SCD++ on PASLCD [ 5 ] . Online masks are emitted when each frame arrives; retrospective masks re-infer the same frames from the state reached after the last frame. The final state suppresses spurious detections that the evidence available online could not yet rule out.
Fig. 6: O-SCD++ with and without predictive photometric uncertainty [ 13 ] on PASLCD [ 5 ] . The rendered uncertainty (blue: low, red: high) is high where the reference map is poorly reconstructed; attenuating the cue there removes false positives caused by rendering artifacts [ 7 ] .
Without
With uncertainty
Dataset
Method
mIoU
F1
mIoU
F1
PASLCD
O-SCD [ 10 ]
0.489
0.643
0.493
0.648
O-SCD++
0.575
0.715
0.602
0.741
CL-Splats
O-SCD [ 10 ]
0.657
0.773
0.688
0.796
O-SCD++
0.703
0.807
0.738
0.831
ChangeSim
O-SCD [ 10 ]
0.404
0.535
0.419
0.553
TABLE VII: Effect of predictive photometric uncertainty [ 13 ] guidance results on PASLCD [ 5 ] , CL-Splats [ 52 ] and ChangeSim [ 11 ] . The rendered uncertainty attenuates the change cue where the reference scene rendering is unreliable, improving the change detection accuracy.
O-SCD [ 10 ]
O-SCD++
Module
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%
TABLE VIII: Per-frame processing time of each module on PASLCD [ 5 ] , in milliseconds and as a share of the frame, measured on an NVIDIA RTX 4090 at a resolution of about 1008×567 (scene-dependent).
Fig. 7: Sparsification of the online masks of O-SCD (left) and O-SCD++ (right) on PASLCD [ 5 ] . In each frame, a growing fraction of pixels is withheld in the order given, and mIoU is computed on the pixels that remain. For O-SCD, pixels are ranked by their distance to its fixed cutoff of 0.5 ; for O-SCD++, by the decision margin of Eq. 22 . Smaller margins and larger uncertainties are withheld first. AUSE is the area between a curve and the oracle, normalized by the plotted range, so that it equals the mean gap to the oracle; lower is better.
Method
Setting
mIoU
Mask agreement
O-SCD [ 10 ]
online
0.4968±0.0057
0.674
O-SCD [ 10 ]
retrospective
0.5089±0.0019
0.853
O-SCD++
online
0.5767±0.0007
0.883
O-SCD++
retrospective
0.5884±0.0000
0.999
TABLE IX: Sensitivity to the order of the inference frames on PASLCD [ 5 ] , over the incoming order and three random permutations. mIoU is the mean ± standard deviation over the four orderings; mask agreement is the mean per-frame IoU between the masks obtained under each permutation and under the incoming order.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Method
LF
Setting
mIoU
F1
CYWS2D [ 4 ]
–
offline
0.273
0.398
VSCD [ 40 ]
–
offline
0.112
0.186
GeSCF [ 23 ]
✓
offline
0.477
0.611
SplatPose+ [ 45 ]
✓
offline
0.237
0.358
SCAR-3D [ 9 ]
✓
offline
0.191
0.289
3DGS-CD [ 6 ]
✓
offline
0.209
0.339
Appendix
TABLE S-I: Offline SCD methods and O-SCD++ on PASLCD [ 5 ] , averaged over all 20 instances. Offline methods use the complete post-change capture. O-SCD++ is causal in its online setting. LF: label-free.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Time (s) ↓
3DGS [ 12 ]
22.21
0.756
0.243
550
3DGS-LM [ 72 ]
22.26
0.756
0.242
340
Speedy-Splat [ 71 ]
22.25
0.760
0.262
399
CLNeRF [ 73 ]
22.27
0.624
0.391
451
O-SCD [ 10 ]
23.82
0.790
0.242
49
O-SCD++
23.82
0.792
0.237
43
Appendix
TABLE S-II: Representation update on PASLCD [ 5 ] , evaluated on novel post-change views and averaged over all 20 instances. Results of the first four methods are from [ 10 ] . O-SCD and O-SCD++ use the same update procedure, differ only in the change masks, and are timed end to end with the same script, including mask generation.
Method
Matcher
Ref. points
Posed
mIoU
F1
O-SCD [ 10 ]
XFeat+MNN
triangulated
89%
0.404
0.535
O-SCD [ 10 ]
ALIKED+LG
triangulated
91%
0.435
0.572
O-SCD++
XFeat+MNN
triangulated
89%
0.474
0.605
O-SCD++
ALIKED+LG
triangulated
91%
0.493
0.626
O-SCD [ 10 ]
XFeat+MNN
rendered depth
100%
0.486
0.638
O-SCD [ 10 ]
ALIKED+LG
rendered depth
100%
0.496
0.650
Appendix
TABLE S-III: Pose front end on the synthetic ChangeSim [ 11 ] , averaged over the test warehouses. Reference points are triangulated from matches between reference frames, as in O-SCD, or lifted with depth rendered from the reference map. Posed is the share of inference images that are registered. Inference images that are not registered receive an empty mask, which scores zero IoU and lowers the overall accuracy. The XFeat+MNN rows with triangulated points are those of Table II . MNN: mutual nearest-neighbour matching. LG: LightGlue [ 68 ] .
Selection
Keyframes
mIoU
F1
Speed-up
All (default)
100%
0.575
0.715
–
Uniform
50%
0.574
0.715
+9%
Voxel coverage
58%
0.568
0.708
+6%
Frustum coverage
16%
0.549
0.686
+12%
Pose-space sampling
43%
0.573
0.714
+9%
Appendix
TABLE S-IV: Keyframe selection for retrieval on PASLCD [ 5 ] , averaged over all 20 instances. Keyframes is the share of reference frames selected as keyframes, which retrieval searches. Speed-up is the throughput gain over searching all reference frames, measured in one session. All other settings are those of O-SCD++.
Online scene change detection requires integrating incoming observations to determine which regions have changed relative to a reference scene. We update change beliefs as new images arrive using generalised Bayesian evidence fusion. However, the prior probability of change and evidence models are unknown beforehand, and numerical Bayesian inference can be costly. We address these challenges by learning offline from automatic edits to a reference 3D Gaussian Splatting (3DGS) model. These edits generate labelled training views without manual annotations. Together with unchanged reference observations, they provide supervision for learning scene-specific evidence models and estimating the effective pixel prior and response statistics. Online prediction combines current evidence with historical support, while closed-form local Bayes factors, parameterized by the learned response statistics, update an accumulated 3D Gaussian field. Object-guided refinement produces the final change masks. Our method achieves state-of-the-art accuracy on PASLCD and CCS3D, with 67.64% and 75.82% mIoU, respectively. On PASLCD, the complete pipeline runs at 17.2,FPS over complete query sequences, the fastest among the evaluated online methods. Project page: https://changruizhu96.github.io/work/lego-scd/
Autonomous robots are increasingly deployed on long duration single- and multi-session missions in dynamic environments, where the ability to identify environmental changes such as fallen trees or opened doors provides important contextual information for online planning. We propose a framework called Change Detection via Scene Decomposition (CDSD) for accurate online geometric change detection using LiDAR or RGB-D sensors. Recent advances in geometric SLAM have made it possible to generate dense, tightly aligned maps without post processing, but comparing global maps across entire sessions is computationally expensive and does not allow for single-session online change detection. CDSD instead spatially decomposes mapped environments into unique scenes where changes can be found efficiently by comparing dense, local subsets of the global map called submaps. As the first submap-based approach for geometric change detection, we identify and address the following core challenges: 1) identifying appropriate scenes for change detection that require minimal redundant information; 2) generating dense and representative submaps for each scene; 3) detecting changes between submaps with differing fields of view; and 4) processing detected changes for real-time map reconstruction. Results demonstrate our algorithm on custom datasets collected at the Army Research Laboratory facility in Graces Quarters, Maryland, and on open-source multi-session change detection datasets.
David Thorne, Samuel Jia Cong Chua, Nakul Joshi +4
University of California, Los Angeles, Los Angeles, CA, USA · DEVCOM Army Research Laboratory (ARL), Adelphi, MD, USA.
Detecting what has changed in an environment is essential for long-term autonomy, yet most change detection settings assume fixed viewpoints, mild misalignment, or only a few changed objects. We introduce Video-based Scene Change Detection (VSCD), which predicts a pixel-wise change mask for each query frame, given a reference and a query RGB video of the same indoor space recorded at different times under unconstrained camera motion. The two videos are not temporally synchronized, and many object instances may appear or disappear. To study this setting, we build a large-scale benchmark with over 1.1 million frames annotated with pixel-accurate change masks, together with a real-world test set for evaluating transfer beyond simulation. We propose a query-centric multi-reference model that learns temporal matching implicitly from change-mask supervision, aligns candidate reference features to the query via local patch correspondence, and fuses per-candidate change features using frame-level and patch-level confidence before decoding a high-resolution mask once per frame. Our approach achieves state-of-the-art performance against strong image- and video-based baselines, and we validate its real-world impact by deploying it on a mobile robot for two downstream applications -- visual surveillance and object incremental learning.
Jiae Yoon, Ue-Hwan Kim
Department of AI Convergence, Gwangju Institute of Science and Technology, Gwangju, Republic of Korea · GIST InnoCORE AI-Nano Convergence Institute for Early Detection of Neurodegenerative Diseases, Gwangju Institute of Science and Technology, 61005 Gwangju, Republic of Korea.