Recent advances in deep monocular visual Simultaneous Localization and Mapping (SLAM) have achieved impressive accuracy and dense reconstruction capabilities, yet their robustness to scale inconsistency in large-scale indoor environments remains largely unexplored. Existing benchmarks are limited to room-scale or structurally simple settings, leaving critical issues of intra-session scale drift and inter-session scale ambiguity insufficiently addressed. To fill this gap, we introduce the ScaleMaster Dataset, the first benchmark explicitly designed to evaluate scale consistency under challenging scenarios such as multi-floor structures, long trajectories, repetitive views, and low-texture regions. We systematically analyze the vulnerability of state-of-the-art deep monocular visual SLAM systems to scale inconsistency, providing both quantitative and qualitative evaluations. Crucially, our analysis extends beyond traditional trajectory metrics to include a direct map-to-map quality assessment using metrics like Chamfer distance against high-fidelity 3D ground truth. Our results reveal that while recent deep monocular visual SLAM systems demonstrate strong performance on existing benchmarks, they suffer from severe scale-related failures in realistic, large-scale indoor environments. By releasing the ScaleMaster dataset and baseline results, we aim to establish a foundation for future research toward developing scale-consistent and reliable visual SLAM systems.
Figures & tables
Fig. 1: Representative example of scale inconsistency. During Sim(3) pose-graph optimization, the SLAM trajectory experiences a sudden scale explosion (highlighted), resulting in an abnormally enlarged reconstruction that diverges from the previously estimated map.
Fig. 2: A diagram illustrating the components of the scale estimation problem.
Dataset
Sequences
Avg. Length
Image Resolution
Multi-floor & Elevation
Pure Rotation
Complex Indoor
Scale Study Feasibility
EuRoC
11
81.2 m
752×480
△
X
X
X
TUM-RGBD
∼ 39
12.2 m
640×480
X
X
X
X
7-Scenes
7
64.3 m
640×480
X
X
X
X
ARKitScenes
>5,000
<100 m
1920×1440
X
O
X
X
Ours (ScaleMaster dataset)
25
152.2 m
1920×1440
O
O
O
O
TABLE I: Comparison of public benchmark datasets with ours.
Fig. 3: A visual comparison of trajectory scales between our proposed ScaleMaster dataset and existing standard benchmarks. This distinct contrast visually demonstrates why existing room-scale benchmarks are insufficient for evaluating long-term scale consistency failures.
Sequence Name
Frames
Path Length (m)
Duration (s)
Environment
Tags
Basement_01
2036
29.11
67
Basement area followed by a staircase ascent.
Indoor, Short trajectory, 3D Map
HotelRoom_01
4217
29.07
141
Interior traversal of a hotel room.
Indoor, Repetitive view, Short trajectory
Lab_01
2167
25.85
72
In-place rotations inside a lab room.
Indoor, Repetitive view, Short trajectory
LargeHall_01
22830
884.12
761
Full loop covering the entire LargeHall.
Indoor, Very long trajectory
LargeHall_02
6576
241.89
219
Evening traversal of a large open hall, low-light.
Indoor, Long trajectory, 3D Map
LargeHall_03
2764
109.89
92
Short loop inside the LargeHall at night.
Indoor, Low-texture risk, Medium trajectory
TABLE II: Summary of collected sequences and their characteristics.
Fig. 4: Our overall experimental pipeline, illustrating the process from data acquisition with our custom rig (left), through ground truth map generation and SLAM processing (center), to the final map-to-map error calculation (right).
TABLE IV: Absolute Trajectory Error (unit: meter) on our dataset.
Sequence Name
Threshold
Chamfer Distance(m)
Drop Rate (%)
Analysis
Library_01
1
0.36
42.5
Severe Failure
10
6.43
0.9
Significant map distortion.
Library_06
1
0.08
1.1
Success
10
0.10
0.0
High-fidelity reconstruction.
Library_07
1
0.52
89.1
Catastrophic Failure
10
9.99
0.0
Near-total map collapse.
TABLE V: 3D Reconstruction Quality of MASt3R-SLAM
Fig. 5: Qualitative comparison of 3D reconstruction of the Library_06 sequence. (a) Ground truth point cloud from LiDAR, color-coded by height. (b) The map reconstructed by MASt3R-SLAM. (c) The alignment of the MASt3R-SLAM map onto the ground truth. (d) Point-to-point distance error visualization, where warmer colors (red) indicate larger geometric inconsistencies between the two maps
Fig. 6: Qualitative comparison on Library_07 demonstrating that low trajectory error does not guarantee correct map scale. (a) The LiDAR ground-truth and (b) the MASt3R-SLAM reconstruction. (c) The overlay and (d) top-down cross-section are generated after applying the single global Sim(3) transformation calculated to align the camera trajectories.
Fig. 7: Short-term scale inconsistency of MASt3R-SLAM on the Station_01 sequence. The repetitive structure and vertical motion cause a severely distorted and overlapping 3D map. This demonstrates the system’s vulnerability even in short, challenging trajectories.
Fig. 8: A long-term optimization failure case for MASt3R-SLAM on the Parking_01 sequence.
Fig. 9: A failure case of inter-session scale ambiguity in MASt3R-SLAM. The image shows the result of processing a single sequence as three independent sessions. The resulting map fragments cannot be aligned into a consistent map as each was generated with a different, inconsistent scale.