MR.ScaleMaster: Scale-Consistent Collaborative Mapping from Crowd-Sourced Monocular Videos

Hyoseok Ju; Giseop Kim

cs.ROApr 13, 2026

MR.ScaleMaster: Scale-Consistent Collaborative Mapping from Crowd-Sourced Monocular Videos

Authors: Hyoseok Ju, Giseop Kim

Organizations: Department of Robotics and Mechatronics Engineering, DGIST, Daegu, Republic of Korea.

Abstract

Crowd-sourced cooperative mapping combines monocular sessions from different front-ends, each an independently reconstructed keyframe sequence. Each session has its own local coordinate frame and may follow an incompatible scale convention. We present MR.ScaleMaster, a backend that accepts image, Sim(3) pose, and point-map packets without requiring a common reconstruction model, camera intrinsics, or front-end-provided metric scale. Our Cross-Front-End Loop Factor (CFL) uses a shared matcher for image correspondences but retrieves matched 3D points from the input point maps, so its scale estimates the inter-session ratio. Scale Preconditioning (SPC) initializes session scales for a Sim(3) anchor-node graph, which then corrects residual scale and drift. For front-ends exposing an incremental scale trajectory, an agent-side Scale Collapse Alarm (SCA) rejects or rolls back false intra-session loops that would otherwise collapse the session scale. We evaluate seven front-ends on KITTI and five on CODa. Based on ground-truth path lengths, session scales in our heterogeneous KITTI setting differ by 57-103x. CFL reduces mean ATE by 37% and inter-session scale error by 60% relative to loop factors from independent pairwise reconstructions. Adding SPC raises the reductions to 74% and 77%, respectively. On CODa, five-session fusion improves mean ATE over single-session runs for all five front-ends, while joint fusion registers 15 sessions from three different front-ends in a single map. Code will be released.

Figures & tables

Fig. 2 : System overview of MR.ScaleMaster. Agent-side SCA rejects or rolls back false intra-session loops. The server applies SPC to initialize session scales, establishes cross-session correspondences, retrieves matched 3D points from the input point maps, constructs

\mathrm{Sim}(3)

CFL s, and optimizes them with odometry in an anchor-node graph.

System class	Agent-side reconstruction assumption	Scale treatment
Metric cooperative SLAM [ 3 , 4 , 5 ]	metric local estimates from visual–inertial, stereo, or range sensing	shared metric scale
Shared-pipeline monocular collaboration [ 16 , 17 , 18 ]	common learned reconstruction pipeline across agents	monocular gauges coordinated within the common pipeline
MR.ScaleMaster	heterogeneous front-ends via session-local packets	independent session scale conventions aligned during fusion

TABLE I : Agent-side reconstruction assumptions and scale treatment in representative cooperative mapping systems.

Symbol	Scale	Quantity
$I_{i}^{r}$	–	keyframe image
$\mathbf{X}_{i}^{r}$	$\lambda_{r}$	session-local pose
$\mathbf{P}_{i}^{r}$	$\lambda_{r}$	per-pixel point map

TABLE II : Keyframe packet provided by session

r

Fig. 3 : Incremental $\mathrm{Sim}(3)$ anchor alignment for three sessions (green, blue, and yellow; orange: anchors; $\mathbf{S}_{1}=\mathbf{I}$ ). (a) Unaligned local maps. (b–c) Inter-session loop closures progressively align the sessions. (d) Full $\mathrm{Sim}(3)$ optimization yields the fused map.

Fig. 4 : Scale-aware backend ablation on KITTI 00 with 15 sessions. (a) $\mathrm{SE}(3)$ optimization absorbs inter-session scale inconsistency as boundary distortions. (b) Anchor-only $\mathrm{Sim}(3)$ aligns sessions while preserving local trajectories. (c) Full $\mathrm{Sim}(3)$ optimization jointly refines anchors and keyframes. (d) Ground truth. Quantitative results are reported in Table III .

Fig. 5 : SCA on a ScaleMaster segment with perceptual aliasing. Without the alarm, an aliasing-induced false loop collapses the MASt3R-SLAM session scale ( $s\approx 0.04$ ); with the alarm, the false loop is rejected and the genuine loop is accepted. Bottom: resulting maps without and with the alarm.

Configuration	ATE RMSE [m]
$\mathrm{SE}(3)$ backend opt.	137.26
$\mathrm{Sim}(3)$ anchor-only opt.	18.92
$\mathrm{Sim}(3)$ full opt.	16.61

TABLE III : MASt3R-SLAM backend ablation on KITTI 00 with 15 sessions.

Fig. 6 : Qualitative multi-session fusion. LingBot-Map reconstructions on KITTI 02 , 05 , and 07 (left to right) contain 15 aligned sessions. KITTI 00 (right) combines five sessions each from MASt3R-SLAM, $\pi^{3}$ , and VGGT-SLAM 2.0; colors denote front-end type.

	ATE RMSE [m] $\downarrow$			$E_{s}$ $\downarrow$
Seq	(A) RayOpt	(B) CFL	(C) CFL $+$ SPC	(A)	(B)	(C)
KITTI 00	134.78	25.05	19.43	1.285	0.153	0.133
KITTI 02	296.39	280.43	78.32	1.284	0.614	0.228
KITTI 05	56.22	18.52	9.79	0.696	0.369	0.112
KITTI 07	21.20	6.39	10.85	0.808	0.136	0.218
KITTI 08	247.58	146.45	80.71	1.343	0.872	0.541

TABLE IV : Heterogeneous KITTI ablation with five sessions each from MASt3R-SLAM,

\pi^{3}

, and VGGT-SLAM 2.0; GT-derived path-length scale ratios span

57

–

103\times

. ATE RMSE [m] and

E_{s}

compare (A) RayOpt, the MASt3R-SLAM ray-and-distance relative-pose construction described in its Eqs. (6)–(7) [ 27 ] , (B) CFL , and (C) CFL with SPC . Lower is better.

Front-end	00	02	05	07	08	Mean
MASt3R-SLAM	$-1.60$	$-5.83$	$+1.23$	$-2.60$	$+26.99$	$+3.64$
$\pi^{3}$	$\mathbf{-19.46}$	$\mathbf{-66.15}$	$\mathbf{-12.71}$	$\mathbf{-16.72}$	$\mathbf{-19.91}$	$\mathbf{-26.99}$
CUT3R	$-4.95$	$\mathbf{-76.10}$	$-3.73$	$-3.62$	$\mathbf{-12.82}$	$\mathbf{-20.24}$
TTT3R	$-1.46$	$\mathbf{-125.04}$	$\mathbf{-10.27}$	$-4.69$	$\mathbf{-19.63}$	$\mathbf{-32.22}$
VGGT-SLAM 2.0	$\mathbf{-10.73}$	$+26.59$	$+8.70$	$-6.84$	$\mathbf{-63.16}$	$-9.09$
LoGeR	$-4.66$	$\mathbf{-37.67}$	$+2.20$	$-1.74$	$\mathbf{-14.75}$	$\mathbf{-11.32}$

TABLE V : CFL ablation across seven front-ends on KITTI (15 same-front-end sessions per run).

\Delta

ATE [m] is CFL minus RayOpt, which builds factors from MASt3R pairwise point maps; all other components are fixed. Negative values favor CFL ; bold marks

\Delta\mathrm{ATE}\leq-10

Sequence	Length	1	1 (tuned)	5	10	15
KITTI 00	3.7 km	TL	103.07	48.69	19.66	16.61
KITTI 02	5.1 km	TL	158.49	113.36	83.06	88.54
KITTI 05	2.2 km	TL	147.15	13.33	10.82	12.01
KITTI 07	0.7 km	TL	2.04	7.05	9.79	8.67
KITTI 08	3.2 km	TL	88.16	166.68	104.83	85.91
Mean		–	99.78	69.82	45.63	42.35

TABLE VI : ATE RMSE [m] on KITTI [ 36 ] vs. session count using MASt3R-SLAM. The tuned single-session baseline relaxes the 3D-correspondence acceptance threshold; all other settings are unchanged. Best per sequence in bold ; TL denotes tracking loss.

	Sequence Info			Same front-end: five sessions $+$ MR.ScaleMaster / single session					Single-session long-context references
Seq.	$L$ [m]	Dur. [s]	KF	CUT3R [ 9 ]	TTT3R [ 10 ]	VGGT-SLAM 2.0 [ 28 ]	LoGeR [ 11 ]	LingBot-Map [ 12 ]	VGGT-Long [ 13 ]	SwiftVGGT [ 14 ]	VGGT-Align [ 15 ]
0	617.7	821.3	579	35.07 / 77.95	11.96 / 15.19	26.37 / 15.31	2.32 / 9.72	3.30 / 11.39	41.04	68.35	65.93
3	1004.7	1472.8	1180	54.92 / 71.46	10.91 / 37.61	30.23 / 34.83	8.43 / 22.96	8.41 / 27.71	54.75	71.32	65.75
4	620.8	823.7	361	41.26 / 58.37	39.94 / 65.12	34.69 / 18.97	2.34 / 31.93	7.31 / 17.67	72.56	72.16	71.82
6	1142.2	1358.2	1663	52.88 / 110.72	29.50 / 47.73	35.89 / 121.57	6.35 / 59.73	8.04 / 55.91	86.84	105.77	119.08
16	1404.5	1706.5	2325	102.87 / 125.00	50.82 / 103.22	46.11 / 144.57	11.54 / 38.67	13.66 / 77.64	100.73	129.56	142.28

TABLE VII : CODa ATE RMSE [m] on identical ORB-selected keyframes. Paired cells report five-session MR.ScaleMaster / same-front-end single-session results; bold marks the lower value. The last three columns are single-session long-context references.

Cat.	Segment	OFF	ON	$\Delta$
Straight	S1 LargeHall_01 (14600–18600)	5.56	3.61	$-1.95$
	S2 LargeHall_02 (1400–5400)	DIV	8.44	–
	S3 Parking_01 (5000–7500)	2.07	1.66	$-0.41$
	S4 Parking_01 (800–3300)	23.03	0.39	$-22.64$
	S5 Stairs_01 (3200–4700)	11.72	0.43	$-11.29$
	S6 Library_01 (3400–4900)	0.32	0.27	$-0.05$

TABLE VIII : SCA on 22 ScaleMaster segments [ 19 ] . ATE RMSE [m] is reported without (OFF) and with (ON) the alarm;

\Delta=\mathrm{ON}-\mathrm{OFF}

. Straight / Rollback test the two gates, and TP contains genuine revisits. DIV denotes divergence; shading marks DIV recovery or

\Delta\leq-4

Fig. 7 : Joint fusion of CODa 0 (TTT3R), CODa 16 (LoGeR), and CODa 21 (LingBot-Map), five sessions each. (a)–(c) Separate fusion; (d)–(f) joint 15-session fusion, with red boxes marking reduced distortion. (g) Per-sequence ATE and ground-truth alignment scale in meters per reconstructed unit; scale spread indicates cross-front-end gauge consistency. (h) Joint trajectories overlaid on ground truth.

Simultaneous Localization and Mapping

Simultaneous localization and mapping estimates a robot’s motion while building a map of an unknown environment.

See all 197 papers

Honey, I Shrunk the Arc de Triomphe!
Jun 1, 2026Yuanbo Xiangli, Hanyu Chen, Xueqing Tsang +13D Foundation Models Metric Depth Estimation
Cornell University · Shanghai Jiao Tong University
Have We Mastered Scale in Deep Monocular Visual SLAM? The ScaleMaster Dataset and Benchmark
Feb 20, 2026Hyoseok Ju, Bokeon Suh, Giseop KimVisual SLAM Indoor Scene Reconstruction
Department of Robotics and Mechatronics Engineering, DGIST, Daegu, Republic of Korea
CoMo3R-SLAM: Collaborative Monocular Dense SLAM with Learned 3D Reconstruction Priors for Outdoor Multi-Agent Systems
May 28, 2026Zhihao Cao, Qi Shao, Shuhao Zhai +4Simultaneous Localization and Mapping Multi-Robot Systems
ETH Zurich, Switzerland. · Harbin Engineering University, China. · University of Liverpool, United Kingdom. +4

Jun 1, 2026J/K move · Enter open · S save

Honey, I Shrunk the Arc de Triomphe!

Yuanbo Xiangli, Hanyu Chen, Xueqing Tsang, Noah Snavely

Cornell University · Shanghai Jiao Tong University

3D Foundation Models Metric Depth Estimation Depth Estimation Depth Completion Monocular Depth Estimation Vision Foundation Model Adaptation

Open paper AskPDF arXiv

Metric scale monocular geometry estimation has seen significant progress through large-scale data aggregation, yet current foundation models suffer from a persistent ''scale-collapse'' phenomenon: distant landmarks and vast landscapes are metrically underestimated. We hypothesize that this performance gap stems from a training data bottleneck, where existing metric-scale datasets are hardware-constrained to homogenous vehicle-captured LiDAR or short-range indoor scans, or consist of synthetic data that lacks the semantic complexity of the physical world. To bridge this gap, we curate a new metrically-grounded, in-the-wild dataset that we call MetricScenes, gathered from a variety of sources including Internet photo collections and stereo imagery. We estimate camera poses and initial depth maps for each scene using off-the-shelf methods, and recover absolute scale from geo-tagged metadata as well as known stereo camera baselines. We also improve the quality of depth maps derived from MetricScenes via a new two-stage Poisson completion method. Fine-tuning MoGe-2 on our dataset significantly mitigates scale-collapse and achieves superior metric accuracy in unconstrained, open-domain scenes while maintaining state-of-the-art performance on standard benchmarks.

Abstract

Simultaneous Localization and Mapping

Explore similar work

Honey, I Shrunk the Arc de Triomphe!

Have We Mastered Scale in Deep Monocular Visual SLAM? The ScaleMaster Dataset and Benchmark

CoMo3R-SLAM: Collaborative Monocular Dense SLAM with Learned 3D Reconstruction Priors for Outdoor Multi-Agent Systems