Research map / Continual learning / Online and replay-based
Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning
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.
Adaptive VIO couples learned visual correspondence and IMU bias predictors with differentiable visual-inertial bundle adjustment, and fine-tunes them online via self-supervised losses from optimized states on EuRoC and TUM-VI.
Problem
Monocular visual-inertial odometry generalizes poorly across environments and sensor attributes; classic VIO relies on hand-crafted visual association and random-walk IMU bias models, while end-to-end learned VIO is less accurate than classic methods.
Approach
Replace hand-crafted visual association with a learned correspondence predictor (DPVO-style, without hidden state) and random-walk IMU bias with a GRU bias predictor, embed both in a differentiable visual-inertial bundle adjustment, and use the optimized poses/velocities as self-supervised targets to update the networks online during tracking.
Key results
- On EuRoC, joint visual and IMU adaptation reduces ATE versus the pre-trained baseline by over 10% on all sequences (e.g., MH1 0.093 to 0.041 m), reaching an average of 0.07 m in the overall comparison.
- On TUM-VI, the method's average ATE is 0.867 m, behind DM-VIO (0.634 m) but ahead of ROVIO, VINS-Mono and OKVIS on average, with better accuracy on 7 sequences than alternatives other than DM-VIO.
- Ablation: visual adaptation alone and IMU adaptation alone each improve over the random-walk/pretrained baseline, with the visual adaptation providing the larger gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt the visual model online while keeping IMU bias as a random walk (ablation) | Adaptive VIO · EuRoC MH1 | RMSE ATE: 0.093 -> 0.043 | Helped |
| adapt IMU bias online while keeping the visual model pretrained (ablation) | Adaptive VIO · EuRoC MH1 | RMSE ATE: 0.093 -> 0.085 | Helped |
| adapt both the visual model and IMU bias online (ablation) | Adaptive VIO · EuRoC MH1 | RMSE ATE: 0.093 -> 0.041 | Helped |
| adapt the visual model online while keeping IMU bias as a random walk (ablation) | Adaptive VIO · TUM-VI room1 | RMSE ATE: 0.271 -> 0.143 | Helped |
| adapt IMU bias online while keeping the visual model pretrained (ablation) | Adaptive VIO · TUM-VI room1 | RMSE ATE: 0.271 -> 0.249 | Helped |
| adapt both the visual model and IMU bias online (ablation) | Adaptive VIO · TUM-VI room1 | RMSE ATE: 0.271 -> 0.138 | Helped |
| combine visual and IMU online adaptation rather than visual adaptation alone (ablation) | Adaptive VIO · EuRoC MH1 and TUM-VI room1 | RMSE ATE: 0.043 -> 0.041; 0.143 -> 0.138 | Helped |
| use Adaptive VIO instead of DM-VIO | VIO systems · EuRoC average | RMSE ATE (m): 0.06 -> 0.07 | Hurt |
| use Adaptive VIO instead of DM-VIO | VIO systems · TUM-VI average | RMSE ATE (m): 0.634 -> 0.867 | Hurt |
| use Adaptive VIO instead of DM-VIO | VIO systems · TUM-VI sequences | RMSE ATE | Mixed |
Surprising findings
- The IMU bias network contributes less to accuracy than the visual network, though it yields lower drift and variance than random walk bias.
Limitations
- Performance on TUM-VI is behind DM-VIO on average and on several sequences.
- Online continual learning is evaluated on a single sequence per dataset, with generalization to others.
- Hyperparameters such as loss weights are set manually.
Experimental setup
Datasets: TartanAir EuRoC MAV TUM-VI
Benchmarks: EuRoC MAV TUM-VI
Metrics: RMSE ATE
Baselines: DPVO MSCKF OKVIS ROVIO VINS-Mono Kimera Online VIO VI-DSO DM-VIO iSLAM SelfVIO
Builds on: DPVO DROID-SLAM DM-VIO VINS-Mono
Closest papers by meaning
- Leveraging Consistent Spatio-Temporal Correspondence for Robust Visual Odometry (AAAI 2025 · similarity 0.87)
- LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry (CVPR 2024 · similarity 0.87)
- EqNIO: Subequivariant Neural Inertial Odometry (ICLR 2025 · similarity 0.86)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.85)
- Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics (CVPR 2025 · similarity 0.85)
- ZeroVO: Visual Odometry with Minimal Assumptions (CVPR 2025 · 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.