Research map / 3D vision / Depth and stereo
MambaVO: Deep Visual Odometry Based on Sequential Matching Refinement and Training Smoothing
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 MambaVO, a learning-to-optimize deep visual odometry system that uses semi-dense geometric initialization, a Mamba-based sequential matching refinement module, and a trending-aware loss penalty, with a loop-closure extension MambaVO++.
Problem
Deep visual odometry based on learning-to-optimize suffers from unstable initialization, ambiguous inter-frame matching, and gradient-variance-driven training instability in the nested matching/bundle adjustment optimization, hurting pose accuracy and robustness.
Approach
Initialize each new frame with semi-dense matching (EfficientLoFTR) plus PnP using monocular depth, refine pixel matches and weights over a point-frame graph with Mamba blocks carrying a history token, backpropagate through a differentiable BA layer, and balance pose and matching losses with a trend-based penalty.
Key results
- On the TartanAir test split, MambaVO averages 0.14 m ATE without loop closure versus 0.17 for DPVO and 0.18 for V2V; MambaVO++ reaches 0.10 m with loop closure.
- On EuRoC, MambaVO averages 0.094 m ATE versus 0.105 for DPVO and 0.117 for V2V (monocular, no loop closure).
- On EuRoC matching AUC@1°, MambaVO reaches 0.471 versus 0.399 for V2V, and runs at about 22 Hz on a single RTX 3090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MambaVO instead of DPVO | Visual odometry · TartanAir test split | ATE [m]: 0.17 -> 0.14 | Helped |
| Add loop closure and global optimization | MambaVO · KITTI | ATE [m]: 50.21 -> 21.84 | Helped |
| Use MambaVO instead of V2V for matching | Visual odometry matching · EuRoC | AUC@1°: 0.399 -> 0.471 | Helped |
| Add context features to GIM (ablation) | MambaVO · EuRoC | ATE [m]: 0.181 -> 0.094 | Helped |
| Add geometric features to GIM (ablation) | MambaVO · EuRoC | ATE [m]: 0.309 -> 0.094 | Helped |
| Add PnP pose initialization (ablation) | MambaVO · KITTI | ATE [m]: 201.16 -> 50.21 | Helped |
| Add history fusion to GMM (ablation) | MambaVO · EuRoC | ATE [m]: 0.179 -> 0.094 | Helped |
| Add Mamba blocks to GMM (ablation) | MambaVO · EuRoC | ATE [m]: 0.318 -> 0.094 | Helped |
| Add GRU to GMM (ablation) | MambaVO · KITTI | ATE [m]: 190.01 -> 50.21 | Helped |
| Add gradient weighting to TAP (ablation) | MambaVO · EuRoC | ATE [m]: 0.111 -> 0.094 | Helped |
| Add trend-based history balance to TAP (ablation) | MambaVO · EuRoC | ATE [m]: 0.101 -> 0.094 | Helped |
Surprising findings
- Removing the Mamba blocks makes the system fail on KITTI, indicating the sequence module is essential for that sequence set.
Limitations
- Trained only on TartanAir; generalization relies on the learned model transferring to other datasets.
- KITTI errors for MambaVO remain large on some sequences (e.g., 00, 02, 08 above 90 m) without loop closure.
Experimental setup
Datasets: TartanAir
Benchmarks: TartanAir EuRoC TUM-RGBD KITTI
Metrics: ATE AUC@1° AUC@2° AUC@5° AUC@10° GPU memory frame rate
Baselines: DROID-VO DPVO V2V DROID-SLAM DPV-SLAM ORB-SLAM3 DeFlowSLAM TartanVO GO-SLAM COLMAP
Builds on: DPVO DROID-SLAM V2V DPV-SLAM EfficientLoFTR
Closest papers by meaning
- Leveraging Consistent Spatio-Temporal Correspondence for Robust Visual Odometry (AAAI 2025 · similarity 0.88)
- ZeroVO: Visual Odometry with Minimal Assumptions (CVPR 2025 · similarity 0.84)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.84)
- LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry (CVPR 2024 · similarity 0.84)
- Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · similarity 0.84)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.84)
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.