Deep Feature Matching SLAM

A simple SLAM system based on feature matching using Deep Learning.

Next steps

• Pose estimation using DFM
• Include Bundle Adjustment
• Test pose estimation in different datasets
• Include Loop Closure
• Optimize DFM Model to TensorRT

How to Install

1. Clone this repository into a catkin_workspace inside the src folder. Then compile it.

catkin_make

2. Install the requirements as follows:

conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia

Then

pip install tqdm opencv-python~=4.5 numpy scipy 

3. Inside the $ROOT/feature_tracker/scripts/feature_tracker.py, modify the first line to the conda environment executable path.

How to Run

TODO

Performance

Currently it was tested on rgbd_dataset_freiburg1_xyz.bag from: https://vision.in.tum.de/data/datasets/rgbd-dataset/download. The reconstruction you can see in the image below.