SGInit-VO

Installation | Usage | License | Thanks

SGInit-VO is a repository containing scripts to replicate the experiment in the paper, titled "Self-Supervised Geometry-Guided Initialization for Robust Monocular Visual Odometry." by Takayuki Kanai, Igor Vasiljevic, Vitor Guizilini, and Kazuhiro Shintani.

The paper and the project page are available:

• arXiv: https://arxiv.org/abs/2406.00929v1
• Project Page: https://toyotafrc.github.io/SGInit-Proj/

The codebase is originally from DROID-SLAM by Teed et al., and is distributed under the Creative Commons Attribution-NonCommercial (CC-BY-NC) License, following the license of TRI-ML/VIDAR, which is adapted to this repository.

Note that, we cannot guarantee future monitoring and official support while we strive to provide assistance or maintenance.

Official implementation of the visual odometry, as well as from-the-scratch pre-training demo.

Installation

Add submodule libraries.

git submodule update --init --recursive
make noroot-build

Usage

Data Preparation

Appropriately modify the CKPT_MNT, DATA_MNT, etc. at Makefile and run:

xhost +local:root # if needed for GUI
make noroot-interactive

We enter the container, and the following procedures are conducted inside the container hereafter.

Download dataset, models, etc. (for your first time)

Please download the materials to your PC by ./tools/download_ddad.sh, ./tools/download_models.sh, etc.

Replicate Visual Odometry (SG-Init + DROID (L), GIF's Right)

./shells/run.sh

Replicate Ablation Model (ZeroDepth + DROID)

./shells/run_ablation.sh

Replicate Depth Evaluation (ResNet18 from ``L'' model)

python3 thirdparty/vidar/scripts/launch.py configs/papers/sginit/inference_resnet18s.yaml

Running on your own data

Please refer to TUTORIAL.md for more details.

License

This repository is released under the the Creative Commons Attribution-NonCommercial (CC-BY-NC) License.

Thanks

We thank the authors of DROID-SLAM for the publicly available code release.

All files without headers are left unchanged and originate from the original codebase.

Otherwise, we left the header on the source files with its copyright.