MIRAGE: Metric Inference and Reliability-Aware Gaussian Estimation from Feed-Forward Vision

Xinzhe Wang; Changjian Jiang; Kaiwen Song; Xudong Li; Kerui Ren; Ran Yi; Lizhuang Ma; Chunhua Shen

cs.CVOct 8, 2026

MIRAGE: Metric Inference and Reliability-Aware Gaussian Estimation from Feed-Forward Vision

Authors: Xinzhe Wang, Changjian Jiang, Kaiwen Song, Xudong Li, Kerui Ren, Ran Yi, Lizhuang Ma, Chunhua Shen,

Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · The University of Hong Kong · University of Science and Technology of China ·

Abstract

Feed-forward 3D priors enable high-quality reconstruction from image streams but often lack reliable metric scale and suffer from long-term trajectory drift. In contrast, LiDAR--inertial estimation provides metrically accurate motion but lacks the image-space consistency needed for high-fidelity reconstruction. We propose $MIRAGE$ , which anchors feed-forward visual geometry to LiDAR--IMU metric motion and regulates online pose corrections according to correspondence support and cross-modal agreement, jointly enabling accurate metric pose estimation and high-fidelity Gaussian reconstruction. Experiments on Oxford Spires, M2DGR and INS show that MIRAGE achieves accurate metric trajectory estimation, outperforming COLMAP, a widely used but computationally expensive reference for vision-only benchmarks. Motivated by the scarcity of long-horizon monocular benchmarks for large-scale indoor reconstruction, we further build a multimodal data-collection platform and use MIRAGE to establish accurate 6-DoF reference trajectories for a new vision-only benchmark. We benchmark recent monocular large-scale reconstruction methods on this dataset, providing a systematic evaluation of their performance in long-horizon indoor scenes.

Figures & tables

Figure 1: We propose MIRAGE , which anchors feed-forward visual geometry to LiDAR–IMU motion through reliability-aware pose refinement. Left: on Oxford Spires Blenheim 02, MIRAGE achieves lower ATE than COLMAP (3.27 vs. 5.50 cm) and higher rendering PSNR than Gaussian-LIC2 (27.82 vs. 21.28 dB). Right: our synchronized multimodal platform supports a long-horizon indoor benchmark with 10 sequences and MIRAGE-established 6-DoF accurate reference trajectories for evaluating vision-only reconstruction. Project Page: MIRAGE-URL .

Figure 2: Pipeline of MIRAGE. We anchor feed-forward visual geometry to LiDAR–IMU motion for metric pose estimation and streaming Gaussian mapping. Local metric constraints support visual pose refinement, while reliability-aware loop correction controls correction magnitude and selectively applies rotation to balance revisit alignment and local motion consistency. The resulting poses drive a triangle-associated Gaussian map. With poses fixed, offline geometry-aware refinement improves scene structure using image residuals and LiDAR support.

	Online						Offline
Sequence	EllipseLIO	Fast-LIO- SLAM	Fast- LIVO2	Gaussian- LIC2	LingBot- Map	Ours	Voxel- SLAM	COLMAP	GLUEMAP	Ours+HBA
Latency	$0.10$	$0.12$	$0.10$	$0.71$	$0.18$	$0.73$	$0.145$	$1.85$	$10.65$	$0.76$
Oxford Spires
Observatory 01	$5.74$	$6.28$	$3.97$	$6.41\!\times\!10^{4}$	$2.46\!\times\!10^{2}$	$3.43$	$2.46$	$6.66$	$95.79$	$2.09$
Observatory 02	$5.23$	$26.22$	$6.83$	–	$3.57\!\times\!10^{2}$	$3.56$	$2.14$	$7.69$	$75.81$	$2.07$
Keble 02	$6.02$	$37.35$	$3.01\!\times\!10^{5}$	$3.08\!\times\!10^{2}$	$3.27\!\times\!10^{2}$	$3.60$	$3.48$	$4.74$	$3.30\!\times\!10^{3}$	$3.66$

Table 1: Translational ATE RMSE (cm,

\downarrow

). Vision-only and metric methods use Sim(3) and SE(3) alignment, respectively. Red/orange mark the best/second-best per sequence, ranked separately for online and offline methods; – denotes failure.

Figure 3: Qualitative trajectory and rendering comparisons. Left: SE(3)-aligned XY trajectories; green boxes show XZ close-ups with equal axis scales. Right: rendering comparisons.

	Oxford Spires						M2DGR			INS
Method	Observatory 01	Keble 02	Keble 03	Keble 04	Blenheim 01	Blenheim 02	room_01	room_02	room_03	indoor_sq1	indoor_sq3	indoor_sq4
Gaussian-LIC2	17.91 0.678 0.570	24.26 0.810 0.421	17.49 0.678 0.554	18.72 0.680 0.548	21.76 0.738 0.530	21.28 0.744 0.535	25.15 0.809 0.258	26.67 0.829 0.251	24.60 0.770 0.341	30.09 0.893 0.299	29.24 0.885 0.308	30.45 0.896 0.301
Ours	26.26 0.827 0.308	27.53 0.836 0.325	26.80 0.821 0.344	24.33 0.741 0.390	26.40 0.796 0.396	27.82 0.837 0.357	25.44 0.824 0.255	26.76 0.831 0.252	26.60 0.843 0.249	30.03 0.907 0.294	29.58 0.901 0.306	29.34 0.900 0.307
Our trajectory + Offline 3DGS	26.04 0.801 0.389	26.87 0.823 0.366	25.75 0.804 0.400	23.19 0.711 0.476	24.82 0.749 0.507	26.65 0.787 0.470	26.69 0.865 0.212	27.02 0.847 0.231	26.42 0.850 0.248	26.36 0.831 0.451	27.71 0.855 0.414	25.60 0.832 0.453

Table 2: Held-out rendering quality. Offline 3DGS ( Kerbl et al., 2023 ) uses 30k iterations with MIRAGE’s online poses fixed. Cells list PSNR (dB,

\uparrow

), SSIM (

\uparrow

), and LPIPS (

\downarrow

) from top to bottom. Red and orange indicate the best and second-best values per sequence and metric.

(a) Pose estimation				(b) Rendering on M2DGR
Stage	Oxford Spires	M2DGR	INS	Configuration	room_01	room_02	room_03
Tracking	82.363	21.868	43.819	w/o visual †	13.90 / 0.504 / 0.561	13.04 / 0.512 / 0.559	12.89 / 0.489 / 0.577
Local	7.784	1.909	7.055	Streaming	23.28 / 0.746 / 0.339	24.41 / 0.762 / 0.344	24.74 / 0.784 / 0.319
Global	4.529	2.067	5.195	+ Photo. refine	24.96 / 0.811 / 0.272	26.29 / 0.816 / 0.275	26.22 / 0.832 / 0.265
Offline (+HBA)	3.093	1.858	3.513	+ Geo. refine	25.44 / 0.824 / 0.255	26.76 / 0.831 / 0.252	26.60 / 0.843 / 0.249

Table 3: Pose and rendering ablations. (a) Scene-averaged ATE (cm,

\downarrow

) over all evaluated sequences. (b) PSNR (dB,

\uparrow

) / SSIM (

\uparrow

) / LPIPS (

\downarrow

\dagger

: camera poses are obtained solely by interpolating LiDAR–inertial poses and applying calibrated extrinsics; RGB supervision for Gaussian reconstruction is retained.

		ATE (m, $\downarrow$ )						Offline 3DGS PSNR (dB, $\uparrow$ )							Stream PSNR
Sequence	Trajectory length (m)	LingBot- Map	GlueMap	LoGeR	Scal3R	ABot- Recon	MASt3R- SLAM	LingBot- Map	GlueMap	LoGeR	Scal3R	ABot- Recon	MASt3R- SLAM	Ours (reference)	Ours
Laboratory 1	54.57	0.299	0.205	0.927	1.901	0.368	0.491	21.65	26.59	20.74	22.16	23.28	17.54	26.10	27.47
Laboratory 2	61.02	0.430	0.225	0.802	1.102	0.675	0.866	19.93	26.41	19.94	22.11	22.53	14.59	25.02	28.47
Laboratory 3	168.46	9.523	15.851	7.364	1.662	0.930	0.729	18.51	22.60	18.96	21.72	22.06	17.22	23.56	26.74
Dining Area 1	112.69	5.058	15.723	10.616	14.772	0.903	2.168	18.42	22.16	18.76	19.09	21.33	16.16	21.68	26.15
Dining Area 2	64.65	0.316	0.206	1.511	0.350	0.427	0.412	20.11	20.45	19.45	20.50	20.87	16.41	19.84	24.06

Table 4: MIRAGE benchmark. ATE: Sim(3)-aligned translational RMSE against MIRAGE references. Offline 3DGS: held-out PSNR with fixed poses; Ours (reference) uses reference poses. Stream PSNR: MIRAGE after refinement. Red/orange indicate the best/second-best per sequence within the ATE and offline 3DGS groups; – denotes trajectory failure.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition
$\mathrm{W},\mathrm{C},\mathrm{L},\mathrm{B}$	World, camera, LiDAR, and IMU coordinate frames.
$\mathbf{T}^{\mathrm{V}},\mathbf{T}^{\mathrm{M}},\mathbf{T}^{\mathrm{F}}$	Visual camera poses, metric camera-pose anchors, and fused camera poses.
$\mathbf{R}\in\mathrm{SO}(3),\ \mathbf{p}\in\mathbb{R}^{3}$	Rotation and translation components of a pose.
$f;\ j,k$	Current image index; visual keyframe indices.
$i;\ m$	LiDAR–IMU temporal index; image correspondence index.
$\tau_{i},\ a_{i}$	Sensor acquisition time and corresponding pose-estimate availability time, with $a_{i}\geq\tau_{i}$ .

Appendix

Table 5: Notation and coordinate conventions.

Figure 4: Observation-supported loop correction. Visual evidence is evaluated first, with geometric fallback when no visually supported candidate is selected. Rotation corrections require additional geometric support.

Figure 5: Acquisition platform and dataset overview. Left: the synchronized RGB–LiDAR–IMU platform, sensor layout, and calibrated sensor transforms. Right: RGB images, colored LiDAR maps, MIRAGE reference trajectories, and IMU signals from four representative sequences spanning laboratories, dining areas, corridors, and offices.

Figure 6: Six additional MIRAGE benchmark sequences. Complementing Figure 5 , each row shows RGB images and colored LiDAR maps (left), alongside MIRAGE reference trajectories and IMU signals (right). IMU plots show normalized 20-second excerpts; trajectory grids are illustrative and do not indicate metric scale.

Sequence	# RGB Images	# LiDAR Frames	Duration (min)	Trajectory Length (m)
Laboratory 1	4,112	2,056	3.42	54.57
Laboratory 2	4,470	2,235	3.72	61.02
Laboratory 3	13,336	6,668	11.11	168.46
Dining Area 1	4,430	2,215	3.69	112.69
Dining Area 2	2,126	1,063	1.77	64.65
Dining Area 3	11,667	5,834	9.72	181.52

Appendix

Table 6: Per-sequence acquisition statistics of the MIRAGE benchmark.

Figure 7: Additional qualitative comparisons on Oxford Spires. Rows, from top to bottom: Observatory 01, Keble 02, Keble 03, Keble 04, Blenheim 01, and Blenheim 02. Left: SE(3)-aligned XY trajectories, with green boxes and connectors indicating local views in the labeled planes at equal axis scales. Fast-LIVO2 is off-scale on Keble 02. Right: held-out ground-truth images, Gaussian-LIC2, MIRAGE, and offline 3DGS using MIRAGE trajectories.

Figure 8: Additional qualitative comparisons on M2DGR and INS. Rows, from top to bottom: room_01, room_02, room_03, indoor_sq1, indoor_sq3, and indoor_sq4. Columns and green trajectory insets follow Figure 7 . The arrow highlights a wall artifact in the Gaussian-LIC2 reconstruction of indoor_sq4.

	Oxford Spires								M2DGR			INS
Stage	Observatory 01	Observatory 02	Keble 02	Keble 03	Keble 04	Blenheim 01	Blenheim 02	Church 03	room_01	room_02	room_03	indoor sq1	indoor sq2	indoor sq3	indoor sq4
Tracking	69.739	97.263	121.833	77.076	57.736	70.032	149.269	15.960	31.694	20.781	13.128	48.784	63.709	30.351	32.431
Local	4.242	4.695	5.083	7.162	10.725	14.654	14.473	1.239	1.828	1.926	1.971	13.148	5.468	5.172	4.433
Global	3.429	3.559	3.595	7.125	9.611	4.495	3.265	1.154	1.873	2.179	2.148	5.666	5.477	5.198	4.441

Appendix

Table 7: Stage-wise translational ATE RMSE (cm,

\downarrow

), with independent fixed-scale SE(3) alignment on common within-run timestamps. Tracking denotes saved visual inputs to metric refinement; Local and Global denote the no-loop and corrected camera trajectories.

Feed-Forward 3D Reconstruction

Feed-forward 3D reconstruction predicts geometry, cameras, and renderable scene representations directly from images without per-scene optimization.

See all 288 papers

Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes
Jun 29, 2026Xi Li, Linyuan Li, Yan Wu +4Panoramic 3D Reconstruction 3D Reconstruction
Realsee, China · Quanzhou University of Information Engineering, China
VIDAR: Visual-Inertial Dense Alignment and Reconstruction via a Geometric Foundation Model
Jul 19, 2026Diyari Mohammed Salih, Lingxiang Hu, Naima AitOufroukh-Mammar +1Monocular 3D Reconstruction
IBISC Laboratory, Université d’Évry Paris-Saclay, Université Paris-Saclay, Évry-Courcouronnes, France.
OmniPoint: Universal Monocular Metric Pointcloud from Any Camera
Sep 8, 2026Botao Ye, Marc Pollefeys, Ming-Hsuan Yang +1Panoramic 3D Reconstruction 3D Reconstruction
Google DeepMind · ETH Zurich · Work done as an intern at Google DeepMind.

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

Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes

Xi Li, Linyuan Li, Yan Wu, Tong Rao+3

Realsee, China · Quanzhou University of Information Engineering, China

Panoramic 3D Reconstruction 3D Reconstruction Camera Pose Estimation Indoor Scene Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

Metric feed-forward 3D reconstruction for panoramic data remains under-explored due to the lack of large-scale panoramic RGB-D training data. We present Realsee3D, a hybrid dataset of 10K indoor scenes (1K real, 9K synthetic) with 299K panoramic viewpoints and precise metric annotations, and Argus, a feed-forward network trained on it for metric panoramic 3D reconstruction. In the sparse unordered capture setting of Realsee3D, a poorly chosen coordinate anchor can cause global pose drift. Argus addresses this with a learned covisibility module that selects the geometrically optimal reference view to anchor the metric world frame. To further improve multi-task learning, we decompose the bidirectional pixel-to-world mapping into interpretable sub-steps with per-step supervision and cross-coordinate joint constraints, reinforcing geometric consistency across prediction branches. On the Realsee3D benchmark, Argus achieves state-of-the-art metric performance in camera pose estimation, depth estimation, and point cloud reconstruction. Project page: https://argus-paper.realsee.ai.

Abstract

Feed-Forward 3D Reconstruction

Explore similar work

Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes

VIDAR: Visual-Inertial Dense Alignment and Reconstruction via a Geometric Foundation Model

OmniPoint: Universal Monocular Metric Pointcloud from Any Camera