Hybrid Visual Odometry System
Explored a two-headed visual odometry architecture that combines RAFT optical flow with an ORB-based classical branch and a confidence selector.
Branches
0
RAFT + ORB
Selector
Confidence
Prediction routing
Output
Pose
Trajectory estimation
Problem
Single-method visual odometry can degrade under domain shifts, texture changes, motion blur, and uncertain optical flow estimates.
Challenge
The system needed a practical decision mechanism that could choose between learned flow and classical features without treating either branch as always reliable.
Architecture
How the pieces fit together.
A RAFT branch estimates dense motion, an ORB branch computes feature-based pose signals, and a confidence-aware selector chooses the final pose estimate.
Architecture View
System structure and decision flow
RAFT Branch
Dense optical flow and learned motion cues.
ORB Branch
Feature matching and classic pose signal.
Confidence Selector
Routes final pose estimate by branch reliability.
Dataset / Inputs
- Frame sequences with motion cues suitable for comparing dense optical flow, feature correspondences, and pose estimation behavior.
Technical Decisions
- Keep learned and classic branches independent before selection.
- Use branch confidence as a first-class signal rather than a logging artifact.
- Compare motion consistency before committing to a final pose estimate.
Implementation Details
- RAFT estimates dense optical flow between consecutive frames.
- ORB extracts and matches keypoints for a classical pose signal.
- A confidence selector routes the stronger prediction into final pose estimation.
Metrics / Results
- The hybrid structure creates a controlled research path for comparing branch confidence, motion consistency, and final trajectory stability.
Lessons Learned
- Hybrid systems are useful when failure modes are complementary.
- Confidence estimation matters as much as raw branch accuracy.
- Classical CV remains valuable as an interpretable fallback branch.
Future Improvements
- Add trajectory-level smoothing and bundle adjustment experiments.
- Benchmark selection logic on sequences with motion blur and low texture.
- Calibrate branch confidence with held-out scene conditions.