Research map / 3D vision / Depth and stereo
Multi-Session SLAM with Differentiable Wide-Baseline Pose Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a monocular multi-session SLAM system built on a RAFT/DPVO-style optical-flow backbone with a differentiable symmetric-epipolar-distance solver for wide-baseline two-view pose and bundle adjustment for visual odometry.
Problem
Estimating camera poses across multiple disjoint monocular videos under a single global reference (multi-session SLAM), where wide-baseline matching and gauge freedoms make existing methods fragile or inaccurate.
Approach
A unified iterative optical-flow backbone that alternates flow updates with a differentiable solver: SED minimization (after weighted 8-point preconditioning) for two-view wide-baseline pose, and bundle adjustment for VO, with Sim(3) trajectory alignment joining sessions.
Key results
- On ScanNet two-view pose, the method reaches 30.9 / 51.1 / 67.6 AUC at 5/10/20 degrees, outperforming SuperGlue, LoFTR, and ASpanFormer (25.6 / 46.0 / 63.3) under the same protocol.
- On EuRoC MAV Vicon 2 sequences, ATE is 0.022 m versus 0.284 m for ORB-SLAM3 (monocular), about 12x lower.
- On ETH3D multi-session, the method succeeds on all 5 groups while ORB-SLAM3 succeeds on only 2 of 5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a shallower correlation pyramid instead of the full pyramid (ablation) | two-view model · Megadepth Val and Scannet Val | Pose error AUC @10°: 53.3 -> 44.3; 33.9 -> 27.1 | Hurt |
| remove attention from the update operator (ablation) | two-view model · Megadepth Val and Scannet Val | Pose error AUC @10°: 53.3 -> 48.6; 33.9 -> 30.5 | Hurt |
| remove ReLU attention from the feature extractor (ablation) | two-view model · Megadepth Val and Scannet Val | Pose error AUC @10°: 53.3 -> 47.4; 33.9 -> 29.7 | Hurt |
| remove pose pre-conditioning (ablation) | two-view model · Megadepth Val and Scannet Val | Pose error AUC @10°: 53.3 -> 35.5; 33.9 -> 20.6 | Hurt |
| remove the SED solver (ablation) | two-view model · Megadepth Val and Scannet Val | Pose error AUC @10°: 53.3 -> 23.8; 33.9 -> 13.1 | Hurt |
| remove the clamping step (ablation) | two-view model · Megadepth Val and Scannet Val | Pose error AUC @10°: 53.3 -> 21.9; 33.9 -> 10.7 | Hurt |
| use iterative optical-flow pose estimation instead of matching followed by a solver | two-view pose estimation · Scannet 1500 test | Pose error AUC @20°: 63.3(65.8) -> 67.6 | Helped |
| use iterative optical-flow pose estimation instead of matching followed by a solver | two-view pose estimation · Megadepth 1500 test | Pose error AUC @20°: 88.9 -> 81.3 | Hurt |
| use the proposed Multi-Session SLAM system instead of mono-visual ORB-SLAM3 | Multi-Session SLAM · EuRoC-MAV V201-203 | RMSE ATE [m]: 0.284 -> 0.022 | Helped |
| use the proposed Multi-Session SLAM system instead of ORB-SLAM3 | Multi-Session SLAM · ETH3D Table | RMSE ATE [m]: 0.018 -> 0.010 | Helped |
Surprising findings
- On MegaDepth, the method is on par with RANSAC-based matchers using OpenCV RANSAC but falls behind when LO-RANSAC is used.
- The SED solver alone, without 8-point preconditioning, converges only when initialized close to the optimum, so the two-stage design is needed.
Limitations
- Two-view accuracy lags RANSAC-based matchers with LO-RANSAC on MegaDepth photo-tourism images.
- The SED solver is non-convex and only converges when initialized near the optimum, requiring the 8-point preconditioning step.
Experimental setup
Datasets: Oxford-Paris 1M Distractors ScanNet MegaDepth TartanAir EuRoC MAV ETH3D
Benchmarks: ScanNet 1500 MegaDepth 1500 EuRoC MAV ETH3D
Metrics: pose error AUC RMSE ATE (m)
Baselines: SuperGlue LightGlue LoFTR MatchFormer ASpanFormer Roessle and Nießner ORB-SLAM3 CCM-SLAM VINS
Builds on: RAFT DPVO DROID-SLAM ORB-SLAM3 Ranftl and Koltun iteratively reweighted 8-point
Closest papers by meaning
- MegaSaM: Accurate, Fast and Robust Structure and Motion from Casual Dynamic Videos (CVPR 2025 · similarity 0.86)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.84)
- MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors (CVPR 2025 · similarity 0.83)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.83)
- PoseCrafter: Extreme Pose Estimation with Hybrid Video Synthesis (NeurIPS 2025 · similarity 0.83)
- 4DTAM: Non-Rigid Tracking and Mapping via Dynamic Surface Gaussians (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.