How swatishirke/visual-inertial-odometry works
Project Type and Purpose
• This is a research‐oriented codebase for Visual-Inertial Odometry (VIO). It implements two parallel pipelines:
a. A classical, tightly-coupled MSCKF (Multi-State Constraint Kalman Filter) VIO (in Code/VIO_Kalman).
b. Deep learning based pose estimation (in Code/Deep_VIO), with separate CNN+LSTM and FlowNet sub-modules for IO (inertial odometry), VO (visual odometry) and full VIO.
File Structure Analysis
Top-level: “Code/” contains two directories plus a root README.
A. Code/VIO_Kalman
• vio.py (entry point)
• msckf.py (EKF/VIO core)
• config.py (parameter definitions)
• dataset.py (EuRoC data loader)
• feature.py, image.py (feature extraction & image handling)
• utils.py, viewer.py (support & visualization)
• test/ (unit tests for feature initialization & utils)
• imgs/ (example output plots)
B. Code/Deep_VIO
• CNN_LSTM/
– Misc/DataUtils.py, Misc/MiscUtils.py (data prep & helper functions)
– Network.py (model definition)
– Train_IO.py, Train_VO.py, Train_VIO.py (training pipelines)
– Test_IO.py, Test_VO.py, Test_VIO.py (inference pipelines)
• Flownet/ (same structure as CNN_LSTM under a different architecture)
• models/ (pretrained .ckpt files under subfolders IO, VO, VIO)
• images/ (example trajectories & result plots)
• report.pdf (detailed network architectures & results)
Key README Insights
• Two main algorithms: EKF-based MSCKF (Python reimplementation of stereo S-MSCKF) and Deep Neural Network based approaches.
• Dataset: EuRoC MAV (stereo images + IMU).
• Dependencies: numpy, scipy, opencv, pangolin (optional viewer).
• Run instructions for each pipeline.
• Results showcased via plots in imgs/.