Skip to content

ScaleMaster Benchmark

Related docs: SLAM Benchmarking Metrics and Datasets, MASt3R-SLAM, DROID-SLAM, SLAM3R / VGGT Foundation SLAM, Bundle Adjustment SLAM, and Nonlinear Solver Diagnostics Crosswalk.

Last updated: 2026-05-23

Executive Summary

ScaleMaster is an ICRA 2026 monocular visual SLAM dataset and benchmark focused on scale consistency. It asks a narrow but important question: do modern deep monocular SLAM systems keep a coherent metric scale when they leave room-scale RGB-D-style settings and enter large indoor, multi-floor, repetitive, low-texture routes?

This is not a visual-inertial benchmark. It is a monocular scale and map-quality benchmark, with ARKit trajectories, RGB images, IMU logs, depth/confidence maps, optimized odometry, and LiDAR reference maps for selected sequences. Its main value for this corpus is the direct map-to-map quality framing: trajectory ATE alone can miss scale collapse, warped floors, and map inconsistency that matter for robots using visual maps for inspection or planning.

What It Contains

The project repository describes:

  • 25 sequences across libraries, large halls, parking and basement areas, stairs, stations, offices, lounges, labs, and hotel-room environments.
  • Large indoor routes, including multi-floor movement, stairs, repetitive views, low texture, low light, and in-place rotations.
  • RGB images from an iPhone 14 Pro at 1920x1440.
  • ARKit VIO odometry, depth maps, confidence maps, and IMU measurements.
  • Optimized odometry from a refinement pipeline using loop-closure verification and GTSAM pose-graph optimization.
  • Seven LiDAR reference maps from a Livox HAP-based reference capture for map-quality evaluation.
  • A map-evaluation script for comparing SLAM reconstructions against reference maps.

The arXiv paper reports that the benchmark evaluates both trajectory accuracy and 3D reconstruction quality, including Chamfer-distance style map-to-map comparison against high-fidelity LiDAR references.

Evaluation Model

Evaluation signalWhat it checksWhy it matters
ATE over long indoor routesGlobal pose error across scale-challenging sequencesShows trajectory drift and gross failures.
Scale consistencyIntra-session scale drift and inter-session ambiguityTargets the main weakness of monocular systems.
Map-to-map qualityReconstruction alignment against LiDAR reference mapsCatches warped or collapsed maps even when some poses look acceptable.
Failure stratificationMulti-floor, repetitive, low-texture, long, and rotation-heavy sequencesMakes the source of scale failure more visible.

Use ScaleMaster beside, not instead of, visual-inertial benchmarks such as LaMAria, Hilti x Trimble 2026, EuRoC, TUM VI, and multi-sensor datasets such as FusionPortableV2.

Domain Fit

DomainFitNote
Deep monocular SLAM researchStrongDirectly targets scale consistency and dense-map quality.
Indoor service robotsConditionalUseful as a visual map-quality stress test, but real robots should add wheel, IMU, LiDAR, fiducial, or floor-plan constraints.
Warehouse / depot / terminal interiorsConditionalRepetitive views and multi-floor routes transfer; handheld capture differs from vehicle-mounted rigs.
Road AVWeakMonocular-only SLAM is not an AV pose backbone.
AirsideWeakUseful only for terminal/hangar visual inspection or monocular fallback studies, not open-apron localization.

Failure Modes It Exposes

  • Monocular scale drift over long routes.
  • Catastrophic map deformation in large halls, libraries, stairwells, and repetitive interiors.
  • False confidence from trajectory-only metrics when dense geometry is wrong.
  • Poor handling of low texture, pure rotations, vertical motion, and loop-heavy routes.
  • Overreliance on learned depth priors without independent metric constraints.

Implementation Notes

  • Treat ScaleMaster as benchmark coverage, not as generally reusable training data, unless access and license terms are confirmed for the target use.
  • Report whether a method uses ARKit, depth, learned priors, IMU, or only monocular RGB.
  • Keep ATE, scale drift, and map-quality metrics separate.
  • Use the LiDAR reference-map subset for geometry checks; do not infer map quality from every sequence.
  • Compare DROID-SLAM, MASt3R-SLAM, VGGT-SLAM-style systems, and classical monocular baselines separately because they use different priors and compute budgets.

Limitations

  • Access is request-gated through the project form.
  • The repository did not expose a clear open-source license during this pass; treat redistribution and commercial reuse as unconfirmed.
  • It is indoor and handheld, not a vehicle-mounted benchmark.
  • It targets monocular scale consistency rather than full multi-sensor operational localization.
  • It should not be used as the only evidence for planner-facing map quality.

Sources

Public research notes collected from public sources.