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

Can Xu; Mingfeng Yuan; Mahan Mohammadi; Steven L. Waslander

cs.ROOct 5, 2026

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

Authors: Can Xu, Mingfeng Yuan, Mahan Mohammadi, Steven L. Waslander

Organizations: University of Toronto Institute for Aerospace Studies and the University of Toronto Robotics Institute, Toronto, ON, Canada

Abstract

Autonomous robots operating over long periods must keep their environmental memory up to date as the world changes between visits. In semi-static environments, an object may be replaced in place by a different but geometrically and semantically similar instance, making the identity change difficult to detect from geometry or coarse semantics alone. We present FOCUS, an uncertainty-aware framework for object-level change detection and map maintenance. We formulate semi-static memory maintenance as probabilistic inference that fuses geometric likelihood with appearance likelihood rendered from a 3D Gaussian map. A recursive three-state estimator maintains whether each mapped object is PERSISTED, REPLACED, or REMOVED. To account for imperfect 3DGS rendering, we model the rendered appearance evidence probabilistically rather than using it as a direct change score, with the model parameters automatically calibrated from a change-free replay of the initial mapping session. We evaluate our method on a new Isaac Sim warehouse benchmark with ambiguous in-place replacements and on the real-world TorWIC dataset. It improves object-level replacement F1 from 0.20 to 0.84 over a probabilistic baseline and transfers to real-world data without manual parameter retuning.

Figures & tables

Fig. 1: System Pipeline. The system first constructs an object-level reference map including a TSDF-based geometric map, and a 3DGS. During later revisits, the system infers each mapped object state by fusing geometry and appearance probabilistic evidence, and updates the map online.

Scene	Method	Level	FP/N (%) $\downarrow$	TP/P ∗	FP/N (%) $\downarrow$	$\mathbf{P}\uparrow$	$\mathbf{R}\uparrow$	$\mathbf{F1}\uparrow$	TP/P ∗	FP/N (%) $\downarrow$	$\mathbf{P}\uparrow$	$\mathbf{R}\uparrow$	$\mathbf{F1}\uparrow$
			Persisted	Replaced					Removed
Small Warehouse	Perceive-Semantix	Voxel	8.2%	6242/37192	8.9%	0.11	0.17	0.13	6692/28926	6.3%	0.15	0.23	0.18
	Perceive-Semantix	Object	33.8%	6/27	23.9%	0.19	0.22	0.20	5/22	26.3%	0.14	0.23	0.18
	DynamicGSG	Voxel	14.5%	15363/37192	12.2%	0.17	0.41	0.25	9170/28926	10.6%	0.13	0.32	0.18
	DynamicGSG	Object	55.9%	15/27	54.1%	0.20	0.56	0.30	14/22	37.7%	0.25	0.64	0.35
	Ours	Voxel	4.7%	27385/37192	1.2%	0.79	0.74	0.76	21073/28926	2.2%	0.62	0.73	0.67

TABLE I: Quantitative change detection results.

Fig. 2: Three examples of in-place replacements (A–C) that preserve category and geometry but differ in wrapping, labels, or color.

Fig. 3: Evidence and object-state belief. An object is observed unchanged in revisit 1 (gray) and replaced by a same-sized box in revisit 2 (red); top frames correspond to the circled observations. Left. In-place replacement. The geometric measurement $\Delta_{t}$ is indistinguishable between the two revisits, whereas the appearance measurement $x_{t}$ clearly separates them. The recursive belief ignores isolated outliers and converges to replaced once appearance evidence accumulates. Right. Real-world removal case.

Fig. 4: Fine-grained change detection and map maintenance in large warehouse. Our method maintains a clean instance-level map, whereas Perceive-Semantix misses most changes and merges adjacent boxes into large semantic segments.

	GT Replaced		GT Removed
Scene	Pred. Rep.	Pred. Rem.	Pred. Rep.	Pred. Rem.
Small Warehouse	25	2	1	21
Large Warehouse	52	4	11	67

TABLE II: State Classification of Detected Changed Objects

Scene	Method	TP/P	FP/N (%) $\downarrow$	$\mathbf{P}\uparrow$	$\mathbf{R}\uparrow$	$\mathbf{F1}\uparrow$
2-1	Perceive-Semantix	9,493/11,378	20.1	0.37	0.83	0.51
	DynamicGSG	9,973/11,378	37.5	0.25	0.88	0.39
	Ours	9,704/11,378	6.0	0.67	0.85	0.75
2-2	Perceive-Semantix	11,691/18,991	16.5	0.52	0.62	0.56
	DynamicGSG	12,415/18,991	23.4	0.45	0.65	0.53
	Ours	13,716/18,991	3.1	0.87	0.72	0.79

TABLE III: Voxel-level removal detection results on TorWIC

Fig. 5: Results on TorWIC 2-4 . Top : instance-level maps colored by object ID, so a color preserved from the reference means the object kept its identity. Ours keeps most identities across the revisit, while Perceive-Semantix reconstructs the scene well but under new IDs. Bottom : detections against the layout ground truth, which show the same pattern.

Scene	Variant	FP/N (%) $\downarrow$	$\mathbf{F1}\uparrow$	R@5% $\uparrow$	FP/N (%) $\downarrow$	$\mathbf{F1}\uparrow$	R@5% $\uparrow$
		Replaced			Removed
Large	Full (closed loop)	4.9	0.70	0.79	10.4	0.51	0.44
Large	Full (map fixed)	4.3	0.72	0.80	7.6	0.58	0.61
Small	Full (closed loop)	1.2	0.76	0.75	2.2	0.67	0.73
Small	Full (map fixed)	1.0	0.77	0.76	1.4	0.72	0.73

TABLE IV: Effect of online map editing on voxel-level performance

Fig. 6: Ablation on the warehouse dataset. Bars show the F1 drop of each variant relative to the full model with the map fixed. Positive values mean the component helps while negative mean removing it improves that revisit.

Explore similar work

OASIS-Map: Object-Level Change Detection in Multi-Session Mapping using Semantic Correspondence Matching
Jul 16, 2026Haedam Oh, Yifu Tao, Nived Chebrolu +1Change Detection Ground-Truth Correspondence
Oxford Robotics Institute, Department of Engineering Science, University of Oxford, UK · Department of Computer Science and Engineering, Indian Institute of Technology Bombay, India
Change-Robust Online Spatial-Semantic Topological Mapping
May 4, 2026Jiaming Wang, Jizhuo Chen, Diwen Liu +4Robot Navigation
National University of Singapore
CubifyGS: Object-Centric 3D Gaussian Splatting for Lifelong Dynamic Scene Maintenance
Jun 27, 2026Bohan Ren, Dianyi Yang, Shiyang Liu +53D Gaussian Scene Reconstruction
School of Automation, Beijing Institute of Technology, Beijing, China · 2Guangzhou Saite Intelligent Technology Co., Ltd.

Jul 16, 2026, cs.ROJ/K move · Enter open · S save

OASIS-Map: Object-Level Change Detection in Multi-Session Mapping using Semantic Correspondence Matching

Haedam Oh, Yifu Tao, Nived Chebrolu, Maurice Fallon

Oxford Robotics Institute, Department of Engineering Science, University of Oxford, UK · Department of Computer Science and Engineering, Indian Institute of Technology Bombay, India

Change Detection Ground-Truth Correspondence Robotic Perception

Open paper AskPDF arXiv

Map representations which are consistent across repeated visits to a real-world semi-static environment are very useful for long-term robotic inspection. In such settings, the scene may evolve while the robot is absent, with objects appearing, disappearing, moving, or being replaced, quickly making a static map outdated. Existing change-detection methods reason through geometry, category-level semantics, or object persistence. However, achieving reliable object association across revisits remains a key challenge, especially under partial views, occlusion, and imperfect segmentation. In this work, we propose OASIS-Map, a multi-session mapping system that maintains a spatio-temporally consistent object-level map by establishing dense patch-level semantic correspondences between temporal observations. These correspondences detect where the scene has changed and incrementally associate objects across revisits as the robot re-observes the environment. We demonstrate OASIS-Map on three challenging real-world scenarios: object rearrangements in 3RScan, visually similar car replacements in a car park, and large-scale scene changes in an outdoor market. We achieve 0.783 F1 on change detection in a car replacement scenario in a car park and 0.667 F1 on moved object association in 3RScan. https://dynamic.robots.ox.ac.uk/projects/oasis-map/

Abstract

Explore similar work

OASIS-Map: Object-Level Change Detection in Multi-Session Mapping using Semantic Correspondence Matching

Change-Robust Online Spatial-Semantic Topological Mapping

CubifyGS: Object-Centric 3D Gaussian Splatting for Lifelong Dynamic Scene Maintenance