ALIKE feature extractor and keypoints detector integrated inside OpenVINS as visual odometry frontend

System requirements

1. x86 CPU
2. CUDA-enabled GPU

Prerequisites

1. Docker desktop or engine
2. Nvidia container toolkit

Installation

1. Clone the project and open it in any IDE with devcontainers support

mkdir openvins && cd openvins
git clone https://github.com/Raspy-Py/OpenVINS-DL-VIO.git .

2. Reopen the project in devcontainer. Preferably, in VSCode
3. Ensure you are in the workspace directory, inside the container

cd /opt/baza

4. Build the project (it takes a while)

# For the first time build you may wanna limit the amout of jobs, if you have less than 20GB of RAM
catkin config --jobs 4 

catkin init
catkin build
source ~/.bashrc # to make all built modules visible 
catkin install

5. Run ALIKE test target to ensure the enviroment is healthy.

/opt/bash/install/alike_extractor/alike_test \
    /opt/baza/src/alike_extractor/models/normal_640x480.onnx \
    /opt/baza/images/test_examples/tum_input.png \
    /opt/baza/images/test_ouputs

Check inside /opt/baza/images/test_examples for such expected results

scores map	labeled scores map	results image

Usage

Run OpenVINS in serial mode

roslaunch ov_msckf serial.launch \
    max_cameras:=1 use_stereo:=false dosave:=true dobag:=true \
    config:=uzhfpv_indoor \
    bag:=/opt/baza/datasets/indoor_forward_3_snapdragon_with_gt.bag \
    path_est:=/opt/baza/evals/uzhfpv_indoor_3_mix.txt

Plot the generated trajectory

rosrun ov_eval plot_trajectories posyaw \
    /opt/baza/src/ov_data/uzhfpv_indoor/indoor_forward_3_snapdragon_with_gt.txt \
    /opt/baza/evals/uzhfpv_indoor_3_mix.txt \
    /opt/baza/evals/uzhfpv_indoor_3_klt.txt

Calculate errors for the generated trajectory

rosrun ov_eval error_comparison posyaw \
    /opt/baza/src/ov_data/uzhfpv_indoor/indoor_forward_3_snapdragon_with_gt.txt \
    /opt/baza/evals/uzhfpv_indoor_3_mix.txt \
    /opt/baza/evals/uzhfpv_indoor_3_klt.txt