O-SCD series
CVPR 2026 Highlight

Changes in Real Time: Online Scene Change Detection with Multi-View Fusion

Chamuditha Jayanga1,2 Jason Lai3 Lloyd Windrim2,4 Donald Dansereau2,3 Niko Suenderhauf1,2 Dimity Miller1,2
1QUT Centre for Robotics 2ARIAM Hub 3ACFR, University of Sydney 4Abyss Solutions

TL;DRDetect what has changed in a mapped scene while you walk through it: no labels, no known poses, multi-view consistent, over 10 FPS, and more accurate than the best offline methods.

Recordings of the O-SCD viewer on three PASLCD scenes, stepping through the inference viewpoints with the detected change masks.
11.2FPS
Online operation on PASLCD, frame by frame
0.486mIoU
Online, already above every offline baseline (best 0.478)
0.552mIoU
Offline with refinement: a new state of the art, in 156 s
8–9×
Faster representation update than re-optimizing from scratch
Abstract

Online, pose-agnostic and label-free

Online Scene Change Detection (SCD) is an extremely challenging problem that requires an agent to detect relevant changes on the fly while observing the scene from unconstrained viewpoints. Existing online SCD methods are significantly less accurate than offline approaches. We present the first online SCD approach that is pose-agnostic, label-free, and ensures multi-view consistency, while operating at over 10 FPS and achieving new state-of-the-art performance, surpassing even the best offline approaches.

Our method introduces a new self-supervised fusion loss to infer scene changes from multiple cues and observations, PnP-based fast pose estimation against the reference scene, and a fast change-guided representation update strategy for the 3D Gaussian Splatting. Extensive experiments on complex real-world datasets demonstrate that our approach outperforms both online and offline baselines.

Accuracy vs. speed

Online, and more accurate than offline

Figure 1. Change-detection F1 against throughput on PASLCD. O-SCD detects changes more reliably than all prior methods, including the strongest offline baselines, at a frame rate comparable to the fastest online approaches. Hover a point for its values; methods below 1 FPS are spread horizontally for legibility.
Method

Overview

Click to open full resolution
Figure 2. We register an incoming inference image Iinfk to the reference representation ℛref with a lightweight PnP-based pose estimator. Using the estimated pose Pinfk and ℛref to render an aligned image Irenk, we extract change cues Ck as a combination of pixel- and feature-level cues. Our self-supervised fusion loss LSSF guides the fusion of all observed change cues into a change representation ℛchange that learns change information from multiple viewpoints and infers change masks Mk. Finally, we selectively reconstruct changed regions to update the 3D Gaussian Splatting representation to the current state ℛinf.
Qualitative results

Comparison with MV3DCD

True positive True negative False negative False positive
Lounge: reference and inference images (top); MV3DCD mask, IoU 0.504 F1 0.670 (bottom left); O-SCD mask, IoU 0.619 F1 0.765 (bottom right).
Figure 3. Top: pre-change and post-change images. Bottom: MV3DCD (left) and O-SCD (right). MV3DCD's hard thresholding and intersection heuristic lead to missed or spurious detections, especially for subtle appearance changes on semantically similar objects (the red-to-blue T-shaped object in Meeting room, the blue-to-black bench in Porch). O-SCD jointly learns complementary change information from pixel- and feature-level cues through its self-supervised loss and captures these fine-grained changes.
Quantitative results

State of the art, online and offline

Scene change detection on PASLCD, averaged over all 20 instances. LF: label-free, PA: pose-agnostic, MV: multi-view consistent. Best in bold, second underlined.

Offline methods Total runtime per sequence
MethodLFPAMVmIoU ↑F1 ↑Time ↓
R-SCD–––0.1180.199194 s
CYWS2D–––0.2730.398189 s
GeSCD✓––0.4770.611298 s
ZeroSCD✓––0.3060.414409 s
3DGS-CD✓✓✓0.2090.339824 s
MV3DCD✓✓✓0.4780.628479 s
O-SCD (ours)✓✓✓0.5520.694156 s
Online methods Operating frame rate
MethodLFPAMVmIoU ↑F1 ↑FPS ↑
ChangeSim–––0.0180.03411.5
CS+CYWS2D–––0.2430.3608.2
CS+GeSCD✓––0.1810.270<1
OmniPoseAD✓✓–0.1680.262<1
SplatPose✓✓–0.1730.281<1
SplatPose+✓✓–0.2370.358<1
O-SCD (ours)✓✓✓0.4860.63811.2
Representation update

Change-guided, efficient Gaussian update

Rendered views from the updated representation compared with CLNeRF and 3DGS trained from scratch; red boxes mark changed regions and yellow boxes unchanged regions.
Figure 4. Rendered views from the updated representation, compared with CLNeRF and 3DGS trained from scratch. O-SCD reconstructs changed regions (red boxes) more accurately while reusing primitives from ℛref to keep unchanged areas (yellow boxes) at high fidelity.
Scene representation update Averaged over all instances and scenes; best in bold, second underlined
MethodPASLCDCL-Splats
PSNR ↑SSIM ↑LPIPS ↓Time (s) ↓PSNR ↑SSIM ↑LPIPS ↓Time (s) ↓
3DGS22.210.75580.242655030.310.93190.1178364
3DGS-LM22.260.75620.242234029.950.93220.1177275
SpeedySplats22.250.76030.261839929.890.93490.1290312
CLNeRF22.270.62390.390745126.290.78670.2235301
O-SCD (ours)23.700.78680.24914230.540.93560.125639

O-SCD matches or exceeds the reconstruction quality of methods that re-optimize the changed scene from scratch, while delivering the updated representation in under a minute, up to 8–9× faster.

Citation

BibTeX

@inproceedings{galappaththige2026online,
  title     = {Changes in Real Time: Online Scene Change Detection with Multi-View Fusion},
  author    = {Galappaththige, Chamuditha Jayanga and Lai, Jason and Windrim, Lloyd and Dansereau, Donald and Suenderhauf, Niko and Miller, Dimity},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {32246--32256},
  year      = {2026}
}

If you use PASLCD, please also cite MV3DCD.

This work was supported by the ARC Research Hub in Intelligent Robotic Systems for Real-Time Asset Management (ARIAM, IH210100030) and Abyss Solutions. C.J., N.S. and D.M. also acknowledge ongoing support from the QUT Centre for Robotics.

QUT Centre for Robotics, ARIAM, The University of Sydney and Abyss Solutions logos