Research map / Detection and segmentation / Tracking
Leveraging Consistent Spatio-Temporal Correspondence for Robust Visual Odometry
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 STVO, a deep visual odometry network that adds temporal propagation of motion states across multi-frame edges and depth-guided spatial attention to a DROID-VO/RAFT-style flow-plus-bundle-adjustment pipeline, improving pose accuracy on four VO benchmarks.
Problem
Hybrid deep visual odometry suffers from noisy and inconsistent two-frame optical flow matching, causing failures in challenging scenes and cumulative drift over long sequences.
Approach
Enforce spatio-temporal consistency in multi-frame flow matching: a Temporal Propagation Module warps per-source motion states across frame-graph edges, and a Spatial Activation Module uses a depth-derived spatial attention matrix to activate context and correlation features before the GRU and differentiable bundle adjustment.
Key results
- On KITTI Odometry, STVO averages 13.56 m ATE vs 22.20 m for DeepV2d (about 38.9% improvement) without KITTI fine-tuning.
- On ETH3D SLAM, STVO averages 0.042 m ATE vs 0.249 m for DPVO (about 83% reduction; abstract claims 77.8% vs previous best).
- On TUM-RGBD, STVO achieves the best average ATE of 0.080 m, versus 0.091 for DPVO and 0.098 for DROID-VO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STVO instead of DPVO | Visual odometry · TUM-RGBD | ATE [m]: 0.091 -> 0.080 | Helped |
| Use STVO instead of DPVO | Visual odometry · EuRoC MAV | ATE: 0.118 -> 0.111 | Helped |
| Use STVO instead of DROID-VO | Visual odometry · ETH3D | ATE [m]: 0.190 -> 0.042 | Helped |
| Use STVO instead of DeepV2d | Visual odometry · KITTI Odometry | ATE [m]: 22.20 -> 13.56 | Helped |
| Add Spatial Activation Module (SAM) (ablation) | DROID-VO · TUM-RGBD | ATE: 0.098 -> 0.089 | Helped |
| Add Temporal Propagation Module (TPM) (ablation) | DROID-VO · TUM-RGBD | ATE: 0.098 -> 0.085 | Helped |
| Add SAM and TPM (ablation) | DROID-VO · TUM-RGBD | ATE: 0.098 -> 0.080 | Helped |
| Replace Depth Anything V2 depth maps with Bundle Adjustment depth (ablation) | STVO · TUM-RGBD | ATE: 0.080 -> 0.082 | Hurt |
Surprising findings
- Ablation shows BA-derived depth alone gives most of the gain of the Spatial Activation Module, suggesting the benefit is not solely from Depth Anything V2 priors.
Limitations
- Full model runs slower (7.3 FPS vs 9.8 for DROID-VO) and uses more GPU memory.
- Depth input relies on Depth Anything V2 or BA-estimated depth.
- DPVO results are re-reported as averages over five runs rather than medians, which changes the comparison from the original paper.
Experimental setup
Datasets: TartanAir
Benchmarks: TUM-RGBD EuRoC MAV ETH3D SLAM KITTI Odometry
Metrics: ATE (m) with scale alignment
Baselines: ORB-SLAM3 DSO DeepV2d TartanVO DROID-VO DPVO
Builds on: DROID-SLAM DROID-VO DPVO RAFT
Closest papers by meaning
- D3VO: Deep Depth, Deep Pose and Deep Uncertainty for Monocular Visual Odometry (CVPR 2020 · similarity 0.89)
- MambaVO: Deep Visual Odometry Based on Sequential Matching Refinement and Training Smoothing (CVPR 2025 · similarity 0.88)
- LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry (CVPR 2024 · similarity 0.87)
- ZeroVO: Visual Odometry with Minimal Assumptions (CVPR 2025 · similarity 0.87)
- Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · similarity 0.87)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 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.