Shiaoming/Python-VO - Githubissues

Python-VO

A simple python implemented frame by frame visual odometry. This project is inspired and based on superpoint-vo and monoVO-python.

We tested handcraft features ORB and SIFT, deep learning based feature SuperPoint, more feature detectors are also possible to be added to this project. For feature matchers, we tested the KNN and FLANN mathers implemented in OpenCV, and the novel deep learning based mather SuperGlue.

Feature detectors

ORB (OpenCV implementation)
SIFT (OpenCV implementation)
SuperPoint

Feature matchers

KNN, FLANN (OpenCV implementation)
SuperGlue

SIFT Keypoints

sift_keypoints

SuperPoint Keypoints

superpoint_keypoints

SIFT+FLANN Matches

sift_flann_match

SuperPoint+FLANN Matches

superpoint_flann_match

SuperPoint+SuperGlue Matches

superpoint_superglue_match

Install

Get this repository

git clone https://github.com/Shiaoming/Python-VO.git
cd Python-VO

Install python packages
```
pip install -r requirements.txt
```

Run

edit dataset path in params/*.yaml;
run python main.py --config params/*.yaml in terminal.

For example, to evaluate the SuperPoint with SuperGlue, run:

python main.py --config params/kitti_superpoint_supergluematch.yaml

Evaluations

Absolute and relative translation errors on KITTI sequence 00 eval

Average relative translation errors on KITTI sequence 00

orb_brutematch	sift_flannmatch	superpoint_flannmatch	superpoint_supergluematch
0.748m	0.085m	0.177m	0.103m

Trajectory of ORB feature with brute matcher on KITTI sequence 00

kitti_orb_brutematch

red: ground truth
green: estimated trajectory

Trajectory of SIFT feature with FLANN matcher on KITTI sequence 00

kitti_sift_flannmatch

red: ground truth
green: estimated trajectory

Trajectory of SuperPoint feature with FLANN matcher on KITTI sequence 00

kitti_superpoint_flannmatch

red: ground truth
green: estimated trajectory

Trajectory of SuperPoint feature with SuperGlue matcher on KITTI sequence 00

kitti_superpoint_supergluematch

red: ground truth
green: estimated trajectory