IRIS-SLAM: Unified Geo-Instance Representations for Robust Semantic Localization and Mapping

Tingyang Xiao; Liu Liu; Wei Feng; Zhengyu Zou; Xiaolin Zhou; Wei Sui; Hao Li; Dingwen Zhang

cs.CVFeb 21, 2026

IRIS-SLAM: Unified Geo-Instance Representations for Robust Semantic Localization and Mapping

Authors: Tingyang Xiao, Liu Liu, Wei Feng, Zhengyu Zou, Xiaolin Zhou, Wei Sui, Hao Li, Dingwen Zhang,

Organizations: Horizon Robotics · D-Robotics · S-Lab, Nanyang Technological University · Hefei Comprehensive National Science Center, Institute of Artificial Intelligence

Abstract

Geometry foundation models have significantly advanced dense geometric SLAM, yet existing systems often lack deep semantic understanding and robust loop closure capabilities. Meanwhile, contemporary semantic mapping approaches are frequently hindered by decoupled architectures and fragile data association. We propose IRIS-SLAM, a novel RGB semantic SLAM system that leverages unified geometric-instance representations derived from an instance-extended foundation model. By extending a geometry foundation model to concurrently predict dense geometry and cross-view consistent instance embeddings, we enable a semantic-synergized association mechanism and instance-guided loop closure detection. Our approach effectively utilizes viewpoint-agnostic semantic anchors to bridge the gap between geometric reconstruction and open-vocabulary mapping. Experimental results demonstrate that IRIS-SLAM significantly outperforms state-of-the-art methods, particularly in map consistency and wide-baseline loop closure reliability.

Figures & tables

Fig. 1 : Pipeline of IRIS-SLAM. The system uses a Unified Geo-Instance Front-end to jointly infer high-fidelity depth and cross-view consistent instance embeddings from monocular RGB streams ( Left ). These outputs are synergized to construct globally consistent dense semantic maps ( Right ). The viewpoint-agnostic instance embeddings act as stable semantic anchors, enabling robust loop closure detection under extreme wide-baselines and challenging perspective changes ( Bottom ).

Fig. 2 : System Architecture of IRIS-SLAM. Our framework establishes a tight coupling between dense geometric reconstruction and instance-level understanding by extending feed-forward 3D foundation models into a shared, multi-view coherent latent space. The pipeline realizes our core contributions through three integrated modules: (1) Unified Geo-Instance Front-end Model : generating consistent geometric-semantic primitives from monocular streams; (2) Instance-Grounded Association : an instance-grounded association mechanism where instance embeddings actively drive data association; and (3) Instance-Guided Loop Closure Back-end : a robust loop closure module leveraging viewpoint-agnostic semantic instance anchors to maintain global map coherence even under extreme pose disparities.

Methods	AVG	360	desk	desk2	floor	plant	room	rpy	teddy	xyz
ORBSLAM3	✗	✗	0.017	0.021	✗	0.034	✗	✗	✗	0.009
Droid-SLAM	0.163	0.202	0.032	0.091	0.064	0.045	0.918	0.056	0.045	0.012
Mast3r-SLAM	0.060	0.070	0.035	0.055	0.056	0.035	0.118	0.041	0.114	0.020
VGGT-SLAM	0.074	0.123	0.040	0.055	0.254	0.022	0.088	0.041	0.032	0.016
VGGT-Long	0.110	0.118	0.058	0.111	0.118	0.071	0.155	0.140	0.120	0.099
Pi-long	0.094	0.115	0.047	0.052	0.160	0.085	0.114	0.143	0.081	0.052

TABLE I : Camera pose estimation comparison with RMSE

\downarrow

on TUM RGB-D [ 36 ] (unit: m). Colors are normalized column-wise from green (best in column) to red (worst in column).

TABLE II : Zero-shot semantic evaluations on ScanNet and Replica datasets. Metrics are normalized column-wise from green (best) to red (worst). IRIS-SLAM achieves competitive results even with estimated (Est/Est) poses and depth.

Panel of Figure 5 — Fig. 3 : Comparative loop closure performance across different overlap thresholds $\tau$ . IRIS-SLAM outperforms baselines with significant margins, especially in wide-baseline scenarios ( $\tau=0.1$ ) where traditional descriptors like NetVLAD and ORB_BoW collapse. Our method maintains a resilient F1-score $\uparrow$ and high Recall@1 $\uparrow$ by leveraging instance and structural consistency.

Method	Precision $\uparrow$		Recall@1 $\uparrow$		F1-score $\uparrow$
Method	$0$ - $30^{\circ}$	$30$ - $60^{\circ}$	$0$ - $30^{\circ}$	$30$ - $60^{\circ}$	$0$ - $30^{\circ}$	$30$ - $60^{\circ}$
NetVLAD	0.219	0.018	0.094	0.007	0.132	0.010
ORB+BoW	0.170	0.051	0.071	0.022	0.100	0.031
SALAD	0.270	0.000	0.130	0.000	0.175	0.000
Ours	0.844	0.297	0.617	0.185	0.713	0.228

TABLE III : Quantitative Evaluation of Cross-View Loop Closure Performance under Increasing Viewpoint Differences.

Fig. 4 : A loop pair with a $67.4^{\circ}$ viewpoint shift. While ours correctly identifies the loop, the best-performing baselines, NetVLAD and SALAD, yield scores only $0.4596$ and $0.2567$ .

Method	Rate	mIoU $\uparrow$	mAcc $\uparrow$	f-mIoU $\uparrow$	f-Acc $\uparrow$
OVO-mapping	10	31.58	45.54	47.43	61.58
Ours (w/o geo)	10	34.75	49.47	48.80	61.46
Ours (full)	10	39.93	55.36	53.62	66.78
	20	36.88	51.77	51.77	65.20
	30	33.46	46.00	48.52	60.23

TABLE IV : Sampling Rate Sensitivity and

S_{\mathrm{geo}}

Ablations Analysis on ScanNet.

Fig. 5 : Qualitative comparison of reconstructed global point clouds on the TUM RGB-D fr1/room sequence: DepthAnything3 (DA3) Stream with Salad loop closure vs. our proposed IRIS-SLAM with instance-guided loop closure.

Fig. 6 : Qualitative comparison of reconstructed global point clouds on the TUM RGB-D fr1/teedy sequence.

Fig. 7 : Qualitative comparison of semantic maps between IRIS-mapping (ours) and OVO-Mapping on ScanNet sequences; distinct colors represent unique object instance IDs. In scenes 0059 and 0181, OVO-Mapping fails to maintain stable instance associations, erroneously merging separate objects, such as foreground items with background cabinets or chairs with tables, into single color-coded entities. Conversely, our joint semantic-geometric tracking achieves precise instance separation. As demonstrated in scene 0000, our method also produces instance masks with significantly sharper boundaries and higher reconstruction fidelity than the baseline.

TABLE V : Zero-shot semantic evaluations on ScanNet and Replica datasets. The best, second best, and third best results in each column are highlighted in pink , orange , and yellow , respectively. Our proposed methods are highlighted in blue.

Table 13

Panel of Figure 14 — Fig. 8 : Visualization of the overlap area calculated via our proposed scheme. The figure presents two sets of concatenated pairs consisting of a query frame and a place frame. The green boundaries delineate the comprehensive co-visible regions identified by our method, with non-overlapping areas dimmed for clarity.

Fig. 9 : Comparison of average Precision and Top-1 metrics across overlap ratio gradients from

0.1

0.5

. The results, averaged across selected ScanNet sequences, demonstrate the significant robustness of our method in wide-baseline scenarios with low visual overlap.

Fig. 10 : Robust loop closure detection under challenging large-viewpoint variations on Replica (top), ScanNet (middle), and TUM (bottom) datasets. By leveraging viewpoint-agnostic semantic anchors, our approach achieves significantly higher loop closure reliability than appearance-based methods under large viewpoint changes.

Fig. 11 : Comparison of loop closure detection results under different overlaps. As the overlap decreases, conventional methods increasingly return incorrect loop closures, whereas our Instance-guided method is able to return correct loop closures across the tested scenarios.

Open-Vocabulary Semantic Mapping

See all 9 papers

GS3LAM: Gaussian Semantic Splatting SLAM
Mar 29, 2026Linfei Li, Lin Zhang, Zhong Wang +13DGS SLAM 3D Gaussian Splatting
School of Software Engineering, Tongji University Shanghai, China · Department of Automation, Shanghai Jiaotong University Shanghai, China
AIM-SLAM: Dense Monocular SLAM via Adaptive and Informative Multi-View Keyframe Prioritization with Foundation Model
Mar 5, 2026Jinwoo Jeon, Dong-Uk Seo, Eungchang Mason Lee +1Simultaneous Localization and Mapping Visual SLAM
The School of Electrical Engineering, KAIST (Korea Advanced Institute of Science and Technology), Daejeon, 34141, Republic of Korea · KAIST InnoCORE LLM, KAIST, Daejeon, 34141, Republic of Korea
VCS-SLAM: Geometry-Validated Semantic Evidence Fusion for 3D Gaussian SLAM
Jun 28, 2026Raman Jha, Shuaihang Yuan, Yi Fang3D Gaussian Splatting 3DGS SLAM
New York University, Electrical & Computer Engineering Dept., Brooklyn, NY 11201, USA. · Embodied AI and Robotics (AIR) Lab, NYU Abu Dhabi, UAE. · New York University Abu Dhabi, Electrical Engineering, Abu Dhabi 129188, UAE.

Mar 29, 2026J/K move · Enter open · S save

GS3LAM: Gaussian Semantic Splatting SLAM

Linfei Li, Lin Zhang, Zhong Wang, Ying Shen

School of Software Engineering, Tongji University Shanghai, China · Department of Automation, Shanghai Jiaotong University Shanghai, China

3DGS SLAM 3D Gaussian Splatting 3D Scene Understanding

Open paper AskPDF arXiv

Recently, the multi-modal fusion of RGB, depth, and semantics has shown great potential in dense Simultaneous Localization and Mapping (SLAM). However, a prerequisite for generating consistent semantic maps is the availability of dense, efficient, and scalable scene representations. Existing semantic SLAM systems based on explicit representations are often limited by resolution and an inability to predict unknown areas. Conversely, implicit representations typically rely on time-consuming ray tracing, failing to meet real-time requirements. Fortunately, 3D Gaussian Splatting (3DGS) has emerged as a promising representation that combines the efficiency of point-based methods with the continuity of geometric structures. To this end, we propose GS3LAM, a Gaussian Semantic Splatting SLAM framework that processes multimodal data to render consistent, dense semantic maps in real-time. GS3LAM models the scene as a Semantic Gaussian Field (SG-Field) and jointly optimizes camera poses and the field via multimodal error constraints. Furthermore, a Depth-adaptive Scale Regularization (DSR) scheme is introduced to resolve misalignments between scale-invariant Gaussians and geometric surfaces. To mitigate catastrophic forgetting, we propose a Random Sampling-based Keyframe Mapping (RSKM) strategy, which demonstrates superior performance over common local covisibility optimization methods. Extensive experiments on benchmark datasets show that GS3LAM achieves increased tracking robustness, superior rendering quality, and enhanced semantic precision compared to state-of-the-art methods. Source code is available at https://github.com/lif314/GS3LAM.

Abstract

Open-Vocabulary Semantic Mapping

Explore similar work

GS3LAM: Gaussian Semantic Splatting SLAM

AIM-SLAM: Dense Monocular SLAM via Adaptive and Informative Multi-View Keyframe Prioritization with Foundation Model

VCS-SLAM: Geometry-Validated Semantic Evidence Fusion for 3D Gaussian SLAM