KinectFusion implemented in Python with PyTorch

This is a lightweight Python implementation of KinectFusion. All the core functions (TSDF volume, frame-to-model tracking, point-to-plane ICP, raycasting, TSDF fusion, etc.) are implemented using pure PyTorch, i.e. no custom CUDA kernels.

Although without any custom CUDA functions, the system could still run at a fairly fast speed: The demo reconstructs the TUM fr1_desk sequence into a 225 x 171 x 111 TSDF volume with 2cm resolution at round 17 FPS with a single RTX-2080 GPU (~1.5 FPS in CPU mode)

Note that this project is mainly for study purpose, and is not fully optimized for accurate camera tracking.

Requirements

The core functionalities were implemented in PyTorch (1.10). Open3D (0.14.0) is used for visualisation. Other important dependancies include:

  • numpy==1.21.2
  • opencv-python==4.5.5
  • imageio==2.14.1
  • scikit-image==0.19.1
  • trimesh==3.9.43

You can create an anaconda environment called kinfu with the required dependencies by running:

conda env create -f environment.yml
conda activate kinfu

Data Preparation

The code was tested on TUM dataset. After downloading the raw sequences, you will need to run the pre-processing script under dataset/. For example:

python dataset/preprocess.py --config configs/fr1_desk.yaml

There are some example config files under configs/ which correspond to different sequences. You need to replace data_root to your own sequence directory before running the script.
After running the script a new directory processed/ will appear under your sequence directory.

Run

After obtaining the processed sequence, you can simply run kinfu.py. For example:

python kinfu.py --config configs/fr1_desk.yaml --save_dir reconstruct/fr1_desk

which will perform the tracking and mapping headlessly and save the results. Or you could run:

python kinfu_gui.py --config configs/fr1_desk.yaml

If you want to visualize the tracking and reconstruction process on-the-fly.

Acknowledgement

Part of the tracking code was borrowed and modified from DeepIC. Also thank Binbin Xu for implementing part of the TSDF volume code which is inspired by Andy Zeng’s tsdf-fusion-python.