Argos: Adapt Rich Geometric Priors for Generalizable Online Scene-Change-Detection

Ruihan Xu; Jiae Yoon; Kaichen Zhou; Ue-Hwan Kim; Luca Carlone

cs.CVOct 7, 2026

Argos: Adapt Rich Geometric Priors for Generalizable Online Scene-Change-Detection

Authors: Ruihan Xu, Jiae Yoon, Kaichen Zhou, Ue-Hwan Kim, Luca Carlone

Organizations: MIT

Abstract

Robots operating in dynamic environments require reliable detection of how their surroundings change over time. Existing learning-based methods largely rely on pairwise 2D image features, which struggle under large viewpoint changes and occlusions, are sensitive to noise, and show limited generalization across domains, while explicit 3D approaches typically require costly offline optimization. We show that the implicit 3D knowledge of Geometric Foundation Models (GFMs) provides a strong basis for addressing these limitations. We introduce Argos, which adapts GFM features for joint scene change detection and 3D reconstruction. To address data scarcity and take a step toward a foundation model for scene change detection, we introduce a large-scale benchmark comprising two synthetic datasets and one real-world dataset, and train jointly across diverse datasets to improve cross-domain generalization. We further introduce Argos-SLAM, a real-time system designed for robotics, which performs online change detection and change-aware 4D mapping. Across benchmarks, our framework substantially outperforms existing baselines, with gains of up to 42.01% in change IoU and 27.91% in F1, while supporting scalable deployment in changing real-world environments.

Figures & tables

Figure 1: We propose Argos , a feedforward model for sub-second change detection and 3D reconstruction, and integrate it into Argos -SLAM for online 4D reconstruction of change-aware point-cloud maps. The figure illustrates how Argos captures scene changes, including the added guitar (3) and removed can (4), while maintaining a consistent spatio-temporal representation across visits.

Figure 2: Attention score visualization of selected layers. The token associated with the added green bottle in $v_{1}$ shows strong attention to the corresponding bottle region across different views within the same visit, while finding no high-attention score in $v_{0}$ , where the bottle is absent.

Figure 3: Argos Architecture. Building on VGGT- $\Omega$ , we introduce a Co-visibility-aware Spatio-Temporal Comparator with three modules: a Camera Reader that injects final camera tokens into intermediate patch features, Inter-session Cross-Attention that aggregates information across visits, and a lightweight DPT Fusion Head that combines multi-layer features to predict per-visit change masks. These masks are then used to disentangle the joint reconstruction into a change-aware map.

Dataset		Pairs	Views	Scenes	Places ∗	Length (s)	Modality	Viewpoint	Trajectory
Released datasets	VL-CMU-CD	1,362	1,362	152	–	–	Real	Identical	-
	PSCD	770	1,540	–	–	–	Real	Identical	-
	ChangeSim	20	130 K	10	–	–	Sim.	Similar	Manual
	VSCD	1,041	1.13 M	210	1.6/1/7	36.0/10.1/116.0	both	Free	Manual
	SceneDiff Dataset	355	172 K	355	1.1/1/2	8.1/2.4/26.1	Real	Free	Manual
Argos -CD (Ours)	Argos -CD-HSSD	501	644 K	167	3.8/1/11	42.8/4.0/203.6	Sim.	Free	Automatic

Table 1: Dataset comparison. Places ∗ and single-visit video length (seconds) are reported as mean / min / max. Pairs counts cross-visit comparisons and Views counts frames with change masks. ∗ Approximated through appearance clustering.

Method		SceneDiff Dataset				Argos -CD-real
Method		Static	Change	mIoU	F1	Static	Change	mIoU	F1
SCD	CSCDNet-finetune	93.33	15.25	54.29	20.23	94.75	8.88	51.81	14.20
	C-3PO ∗ -finetune	94.42	17.88	56.15	22.47	95.49	8.71	52.10	12.41
	GeSCF †	69.69	5.04	37.37	15.12	63.64	2.95	33.30	11.13
	RobustSCD-finetune	94.41	18.97	56.69	21.25	94.46	9.06	51.76	15.92
	ZSSCD †	88.63	6.74	47.69	13.22	79.97	3.79	41.88	11.47

Table 2: Real-world generalization study. Argos performs consistently better than learned baselines even without fine-tuning to the real-world datasets. † indicates zero-shot methods.

Panel of Figure 4 — Figure 4: Qualitative results. We show results of the best-performing models. Argos shows cleaner masks and fewer false positives or false negatives in real-world. Appendix A.7 shows more results.

Method	Argos -CD-HSSD				Argos -CD-SceneSmith				VSCD ∗∗
Method	Static	Change	mIoU	F1	Static	Change	mIoU	F1	Static	Change	mIoU	F1
CSCDNet	92.91	38.10	65.51	36.53	94.32	31.11	62.72	34.83	94.03	31.95	62.99	32.44
C-3PO ∗	91.85	37.65	64.75	37.68	94.46	38.68	66.57	40.77	93.88	32.98	63.43	32.90
GeSCF †	76.73	13.81	45.27	28.41	81.56	13.72	47.64	29.80	77.74	12.26	45.00	23.98
RobustSCD	88.86	33.18	61.02	34.76	92.70	33.92	63.30	36.85	89.54	27.18	58.36	30.59
ZSSCD †	90.03	16.58	53.31	16.22	91.63	15.15	53.39	15.49	93.00	15.72	54.36	16.02

Table 3: Performance on Video Scene Change Detection. ∗ C-3PO architecture disables the “remove” feature branch. ∗∗ The VSCD dataset is modified to contain “appeared” labels in both reference and query frame.

Method	PSCD				ChangeSim ∗				VL-CMU-CD
Method	Static	Change	mIoU	F1	Static	Change	mIoU	F1	Static	Change	mIoU	F1
FC-EF	84.97	37.99	61.48	48.14	66.21	12.44	39.32	20.72	91.36	36.72	64.04	45.61
FC-Siam-diff	89.55	43.72	66.64	52.67	85.77	21.12	53.44	33.61	96.17	55.97	76.07	64.21
FC-Siam-conc	86.96	41.76	64.36	52.13	81.63	19.01	50.32	29.40	94.40	50.81	72.61	58.71*
CSCDNet	92.57	48.84	70.70	60.81	92.77	26.48	59.62	37.64	97.57	69.17	83.37	76.90
DR-TANet	91.17	43.40	67.28	55.79	90.91	27.57	59.24	38.53	97.16	65.81	81.48	73.98

Table 4: Performance on paired-image Scene Change Detection. Argos outperforms all baseline methods, and the foundation variant of Argos scores closely to the dataset specific variants.

Architecture Variant				VSCD**				PSCD
Backbone	CDHead	CrossAttention	CameraReader	Static	Change	mIoU	F1	Static	Change	mIoU	F1
SAM	$\checkmark$	$\checkmark$	$\checkmark$	91.92	27.89	59.90	30.64	92.05	46.08	69.07	58.59
VGGT- $\Omega$	$\checkmark$	$\times$	$\times$	96.34	47.97	72.16	42.52	94.64	63.88	79.27	69.26
VGGT- $\Omega$	$\checkmark$	$\checkmark$	$\times$	97.89	63.51	80.70	50.01	94.48	62.11	78.29	69.08
VGGT- $\Omega$	$\checkmark$	$\times$	$\checkmark$	93.74	25.96	59.85	29.37	94.63	63.53	79.08	69.16
VGGT- $\Omega$	$\checkmark$	$\checkmark$	$\checkmark$	98.24	68.69	83.46	53.14	94.80	64.56	79.68	69.89

Table 5: Architecture Ablation. Inter-session cross attention and Camera reader complement each other to achieve the best performance, while VGGT-

\Omega

backbone provides a strong foundation.

Figure 5: Qualitative result of the online system in a multi-room scene. Argos -SLAM localizes additions and removals across object scales, such as a removed chair, an added honey bottle, and a relocated Husky robot, consistently in both 2D observations and the global 4D map.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: Synthetic dataset trajectory generation process. The steps ensure maximum coverage of visible objects with a small number of viewpoints to ensure sufficient and efficient coverage of the environments.

Setting	#Samples	COLMAP Success	Avg. Query Frames	Full Success (32/32)	$\geq$ 90% Success (29+/32)
32+32	220	141 (64.1%)	7.5 / 32	4 (1.8%)	4 (1.8%)
64+32	220	181 (82.3%)	11.9 / 32	6 (2.7%)	15 (6.8%)
128+32	220	199 (90.5%)	13.8 / 32	7 (3.2%)	15 (6.8%)
256+32	220	195 (88.6%)	13.0 / 32 4	(1.8%)	9 (4.1%)

Appendix

Table 6: O-SCD pose-registration prescreen coverage. Across all 220 classic-VSCD test candidates, increasing the reference-frame budget from 32 to 128 provide enough datapoint for reliable evaluation.

Method	Static	Change	mIoU	F1
O-SCD (R=32)	88.42	19.18	53.80	26.28
O-SCD (R=128)	90.20	28.95	59.57	40.64
VSCDNet	90.22	22.45	56.33	31.05
Ours (VSCD)	93.21	37.70	65.46	49.41

Appendix

Table 7: Comparison against O-SCD on the top-12 classic-VSCD scenes O-SCD can register. When given enough reference frames, O-SCD outperforms the VSCDNet baseline but still falls short of our proposed method.

Figure 8: Qualitative comparison between O-SCD, VSCDNet, and Argos . Even with 128 reference frames, O-SCD produces sensible predictions—particularly over removed objects—but still yields many false positives (FP) and false negatives (FN).

Panel of Figure 9 — Figure 9: O-SCD limitation analysis. Even when COLMAP and Gaussian Splatting succeed, poor rendering quality or incorrect camera registration can still lead to noisy predictions.

Method	ChangeSim
Method	Static	Change	mIoU	F1
FC-EF	65.68	16.58	41.13	27.21
FC-Siam-diff	84.56	25.03	54.80	38.55
FC-Siam-conc	80.63	23.32	51.95	35.44
CSCDNet	90.61	24.05	57.33	35.70
DR-TANet	89.14	28.02	58.59	39.71

Appendix

Table 8: Performance on ChangeSim under original protocol.

Method		Video length			Graphic level		Object changes			Synthetic	Captured by		Real-world
Task	Model	Low	Mid	High	Mid	High	Low	Mid	High	Overall	Robot	Human	Overall
AOD	TCF-LMO	17.1	20.1	22.2	18.7	20.8	15.2	21.0	21.8	19.7	7.4	12.0	10.3
AOD	PBCD-MC	24.6	28.0	26.5	26.3	27.4	23.4	28.0	27.8	26.8	16.8	15.7	16.1
SCD	CSCDNet	18.8	21.4	16.9	21.7	17.5	20.0	20.1	19.0	19.8	8.2	9.6	9.1
	DR-TANet	20.5	22.0	17.2	22.9	17.9	20.8	20.5	20.6	20.6	9.1	13.1	11.6
	C-3PO	22.7	26.1	20.8	30.1	16.9	18.8	25.1	27.8	24.1	8.0	13.9	11.7

Appendix

Table 9: Overall performance on the original VSCD benchmark. We report frame-wise F1 score (%) on the synthetic dataset and the real-world test set, together with results stratified by video length, graphic level, and the number of changed objects.

Method	Number of input images $N$
Method	2	4	8	16	32	64	128
TERDNet	0.076	0.095	0.13	0.26	0.51	1.03	2.05
C-3PO	0.008	0.010	0.015	0.026	0.049	0.10	0.21
VSCDNet	0.030	0.059	0.12	0.25	0.48	0.96	1.91
SceneDiff †	22.8	32.6	51.6	96.6	210	485	OOM
O-SCD ‡	Fail	Fail	Fail	187	209	194	286

Appendix

Table 10: Inference runtime vs. number of input images.

Figure 10: Qualitative result of online robot mapping in a small-size workshop environment.

Figure 11: Qualitative result of online robot mapping in a medium-sized common space area.

Figure 12: Qualitative result of the offline mapping from iPhone-captured video in a multi-room apartment.

Figure 13: Qualitative result of online robot mapping in a small-size office environment.

Panel of Figure 14 — Figure 14: Qualitative results of SCD datasets. SCD contains image pairs with labels only on the $v_{0}$ or $v_{1}$ image. Argos consistently produces predictions with less noise.

Figure 15: Qualitative result of the SceneDiff dataset.

Figure 16: Qualitative result of the Argos -CD-real dataset. Stretched vertically to provide a better view.

Figure 17: Qualitative result of the VSCD dataset. Stretched vertically to provide a better view.

Figure 18: Qualitative result of the Argos -CD-HSSD dataset.

Figure 19: Qualitative result of the Argos -CD-SceneSmith dataset.

Explore similar work

Online Geometric Change Detection via Scene Decomposition
Sep 15, 2026David Thorne, Samuel Jia Cong Chua, Nakul Joshi +4Change Detection Simultaneous Localization And Mapping
University of California, Los Angeles, Los Angeles, CA, USA · DEVCOM Army Research Laboratory (ARL), Adelphi, MD, USA.
VGGT-CD: Training-Free Robust Registration for 3D Change Detection
May 16, 2026Wei Zhang, Songhua Li, Yihang Wu +2Image Registration Change Detection
Northwestern Polytechnical University, Xi'an, China · Northwestern Polytechnical University Xian, 710072, China
FOCUS: Fine-Grained Open-Vocabulary Change Detection for Uncertainty-Aware Semi-Static Scenes
Oct 5, 2026Can Xu, Mingfeng Yuan, Mahan Mohammadi +1Robotic Perception
University of Toronto Institute for Aerospace Studies and the University of Toronto Robotics Institute, Toronto, ON, Canada

Sep 15, 2026, cs.ROJ/K move · Enter open · S save

Online Geometric Change Detection via Scene Decomposition

David Thorne, Samuel Jia Cong Chua, Nakul Joshi, Aiden Wong+3

University of California, Los Angeles, Los Angeles, CA, USA · DEVCOM Army Research Laboratory (ARL), Adelphi, MD, USA.

Change Detection Simultaneous Localization And Mapping Robotic Perception Scene Understanding

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Online Geometric Change Detection via Scene Decomposition

VGGT-CD: Training-Free Robust Registration for 3D Change Detection

FOCUS: Fine-Grained Open-Vocabulary Change Detection for Uncertainty-Aware Semi-Static Scenes