PRISM-SLAM: Probabilistic Ray-Grounded Inference
for Scale-aware Metric SLAM

Trajectory

Verified metric trajectories from PRISM-SLAM (monocular RGB): indoor (TUM) and outdoor (KITTI seq. 03).

System Architecture

DSUG: Temporal Depth Consistency as Uncertainty Proxy

Metric Scale in Monocular SLAM

Experimental Results

All runs use monocular RGB on an NVIDIA RTX 4500 Ada unless noted. We benchmark TUM RGB-D, 7-Scenes, and BONN Dynamic, reporting ATE RMSE in centimeters under standard Sim(3) alignment and, for PRISM-SLAM, additional SE(3) ATE using the system’s own metric scale (no oracle scale correction). Unless noted, numbers are the median of three independent runs.

TUM RGB-D — Static Sequences (fr1)

ATE RMSE (cm). SE(3) uses runtime metric scale (no oracle Sim(3) correction). On fr1/xyz, the SE(3) ATE closely matches the Sim(3) ATE with only 3.3% scale error.

TUM RGB-D — Static & Dynamic Sequences (fr3)

ATE RMSE (cm). Baselines evaluated under Sim(3) (Umeyama); PRISM additionally reports SE(3) under its own runtime metric scale. The same strict metric fidelity extends to fr3 (paper Sec. 4.2): e.g. sit-static Sim(3) 1.6 cm vs. SE(3) 1.8 cm. WildGS-SLAM uses Metric3D v2 but reports only Sim(3) ATE (∼0.5 FPS, GPU-only).

7-Scenes Indoor Localisation

ATE RMSE (cm). KeyNet frontend with 4096 features (N_f = 4096; paper Table 7). Mean Sim(3) ATE is 8.8 cm across all seven scenes. Heads is dominated by extreme rotation and weak metric observability (large SE(3) gap); the Mean column reports Sim(3)/SE(3) averaged over the six non-degenerate scenes, matching the project table convention.

BONN Dynamic — Monocular RGB vs. RGB-D (Sim(3) / SE(3))

ATE RMSE (cm). RGB-D baselines use active hardware depth (standard Sim(3) reporting); dashes mark sequences not evaluated in the paper table. PRISM reports both alignments: on balloon2 and pers trk, SE(3) (17.7 / 39.5 cm) closely tracks Sim(3) (14.0 / 36.7 cm), evidencing consistent runtime metric scale under heavy dynamics. See the paper (Table 6) for DSUG-only ablations on these splits.

BONN — DSUG Ablation (SE(3) ATE, cm)

Strict SE(3) alignment. Disabling DSUG spikes the mean error by +7.1 cm; on pers trk, ATE rises from 39.5 cm to 51.1 cm (paper Table 6).

KITTI Odometry (outdoor, first 500 frames)

Outdoor stress test with large depth range and high ego-speed; PRISM runs monocular RGB with metric SE(3) ATE. ORB-SLAM2 uses stereo to fix scale (reference only).

Sequence 03. Stereo row follows the paper’s Table 9 (†): two-camera input to bypass monocular scale ambiguity.

ViT-Driven Loop Closure (DA3 [CLS])

DSUG-Gated Dense Metric Reconstruction

Loading…

fr1/desk2

Loading…

fr1/room

Offline pipeline (paper Sec. 3.7, Table 10): batches of globally optimized poses feed DA3 multi-view inference (N = 5, GT extrinsics in the evaluation), then DSUG-filtered depth is fused in a TSDF (DA3-MV uses 2 cm voxels vs. 1 cm for the filtered single-view baseline). On fr1/desk2, DA3-MV cuts Chamfer distance to 17.0 cm vs. 23.5 cm and raises F@5 cm to 31.7% (vs. 20.7%) relative to the filtered baseline.
Drag to rotate · scroll to zoom · right-click to pan.

Ablation Study

Paper Table 5: mean ATE (cm) as SE(3) over fr3/sit-static, fr3/walk-static, and fr1/xyz; full system 2.12 cm. Removing the Plücker ray-distance factor (+0.89 cm) is the dominant failure mode, consistent with cross-view ray constraints resolving the scale null-space. Disabling DSUG adds +0.22 cm on these mostly static splits (its impact is much larger on BONN dynamics; Table 6). Log-domain Kalman (+0.16 cm) and confidence-weighted WLS vs. a simple unweighted variant (+0.11 cm) further stabilize asynchronous scale fusion.

BibTeX

@article{prism-slam2026,
  title={PRISM-SLAM: Probabilistic Ray-Grounded Inference for Scale-aware Metric SLAM},
  author={Anonymous},
  journal={Under review},
  year={2026}
}