Pow3R-SLAM: Real-Time RGB-D SLAM with 3D Reconstruction Priors

3 min read

Pow3R-SLAM turns MASt3R-SLAM into an RGB-D system by feeding sensor depth into the two-view network Pow3R as a conditioning input, instead of fusing depth into the map as geometry. Depth enters at three separate places: it conditions the network’s prediction, it sets the metric scale of each pointmap (a clipped geometric mean of per-pixel depth ratios), and it anchors keyframe depths inside the backend solve. Pow3R has no matching head, so the authors rebuild loop retrieval from its encoder tokens and per-match confidence from its confidence head. They also add map-only keyframes to densify the map and a hybrid mode that tracks every second frame with point-to-plane ICP.

Findings

  • Over 24 sequences from TUM, 7-Scenes and Replica on an RTX 4090: 903 s against 1443 s for MASt3R-SLAM (18.6 against 14.0 FPS), 15% lower mean keyframe ATE, and 3.1x lower per-frame ATE with no scale fitted (0.037 m against 0.116 m). The hybrid mode runs at 25.3 FPS.
  • Maps: mean Chamfer 3.1 cm against 4.3 cm, RMSE Chamfer 4.1 cm against 6.9 cm over 15 scenes.
  • Depth conditioning carries most of the gain (mean ATE 0.083 m to 0.032 m in ablation). Keeping only 3% of depth pixels at the conditioning site stays within 2.2 mm of the full result.
  • Dense wrong depth is worse than sparse depth. Multiplicative noise with standard deviation 0.2 raises mean ATE from 0.026 m to 0.197 m; at 0.1, noisy depth still beats no depth.
  • ORB-SLAM3 in RGB-D mode loses 3 of 9 TUM sequences and Pow3R-SLAM completes all of them. On a 16-sequence ETH3D panel it is the other way round: Pow3R-SLAM fails 2 (planar_3, repetitive) while MASt3R-SLAM completes all 16.

Limits

  • The authors say plainly that the depth prior “does not make the estimators more robust”. On self-similar structure (7-Scenes stairs, ETH3D sofa_2) the network predicts confident, complete geometry, and the matcher finds wrong but mutually consistent correspondences.
  • Indoor RGB-D only, one fixed calibration, a desktop GPU (about 41 ms per network pass). Code is promised on acceptance.

Relevance

Underwater, the depth would come from stereo or a learned model, and in turbid water both give dense depth that is confidently wrong: the regime this paper shows hurts most. Repeating texture is common underwater too. The stairs failure shows consistent outliers that a confidence gate cannot catch; see outlier-robust estimation and robust robotics.