DMVOGV

Deep Monocular Visual Odometry for Ground Vehicle

Motivation

Most learning-based visual odometry aim to learn six degree-of-freedom motion despite the motion in training dataset is contrained. We simplify the learning targe by only learning the major motion of a ground vehicle. we contruct a very light model to learn the main motion of ground vehicle which can run in real-time on CPU.

How to run this code

1. install dependency (Supposed you are with a new Ubuntu with nothing installed)

sudo apt-get update
sudo apt-get install python3 python3-pip
python3 -m pip install -r requirements
check whether you have a GPU on your computer
- if yes (you have a GPU with 2G or more graphic memory) you have to install nvidia driver, cuda by yourself;
- according to pytorch offical, install corresponding pytorch and torchvision

2. download training dataset

download kitti odometry dataset http://www.cvlibs.net/datasets/kitti/eval_odometry.php
put image list file in dataset/kitti/

3. train

adjust parameter (batch size, training data, evaluation data and so on) in train.sh refer to src/options.py
sh train.sh
After the training finished, you will got validation path in checkpoint/saved_result and saved_model parameter in checkpoint/saved_model

4. test

adjust parameter (batch size, testing data and model_to_load) in test.sh refer to src/options.py
sh test.sh

5. Code

src/simple_train.py main script of the project, intialize model, optimization method, dataloader and train
src/VONet.py Realization of VO estimation model
src/loss_functions.py Loss function
src/data_loader.py dataloader which only load consequent images, src/data_loader_random.py, dataloader which can load images with random distance to contruct an input image pair
src/options.py options
src/transformation.py motion transformation: eular<->mat<->so3<->quaternion, pose <->motion, line<->mat and so on.

TimingSpace / DMVOGV