A basic implementation of KinectFusion [1] in Python inspired by https://github.com/andyzeng/tsdf-fusion-python.
It's light and fast (25~30 FPS with my RTX3090).
It has been tested on tabletop scenes for robot manipulation and small rooms for navigation :)

• Python >= 3.10
• CuPy (check the installation guide on the official website)
• Open3D (pip install open3d)
• cupoch (pip install cupoch)

git clone [email protected]:shiyoung77/KinectFusion-python.git
python -m KinectFusion-python.main --dataset {dataset_path} --video {video_name}

If you want to save not only the poses and reconstructed point cloud but also the TSDF volume, add --save_tsdf option.
By default it saves the result in the video folder. You can specify the output directory with --output_dir option.
Check main.py for more command line options. Check the kf_config.py for the default parameters for reconstruction.

You should have your dataset in the following format.

{dataset_path}/
    {video_name}/
        color/
            0000-color.png(or jpg)
            0001-color.png
            ...
        depth/
            0000-depth.png
            0001-depth.png
            ...
        config.json
    {second_video_name}/
        ...

config.json should contain the camera information. An example config.json is as follows.

{
    "id": "video0",
    "im_w": 640,
    "im_h": 480,
    "depth_scale": 1000,
    "cam_intr": [
        [
            1066.778,
            0,
            312.9869
        ],
        [
            0,
            1067.487,
            241.3109
        ],
        [
            0,
            0,
            1
        ]
    ]
}

An example RGB-D video capture at Rutgers CS Robotics Lab could be found at this google drive link.

[1] Newcombe, R. A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A. J., ... & Fitzgibbon, A. (2011, October). Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality (pp. 127-136). IEEE.