The render strategy is to combine the 3D Gaussian models in one scene with different resolution to render the final image.

• First, use the original training script with different densify_grad_threshold:

•   python train.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to save model> --densify_grad_threshold 0.1

• Second, use the render_multiModel.py script to render the final image:

•   python render_multiModel.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to trained model> --model_paths <path to other trained models that combined with>

Command line arguments for render_multiModel.py:

• --model_paths: path to other trained models that combined with
• --strategy: render strategy, default is 'dist', other options are 'fov' and 'distFov'
• --render_image: save rendered image

LineGS is a 3D line segment representation on 3D Gaussian splatting. This method is a post-processing step after training the 3D Gaussian splatting model.

This method follows the steps below:

• train the 3D Gaussian splatting model with the original training script:

•   python train.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to save model>

• use Line3D++ or other 3D line segment detection methods to generate initial 3D line segments.
• use line3d.py to implement LineGS, and the result will be saved in the folder named the same as the folder that contains 3D line segments with prefix '_test'.

•   python line3d.py -m <path to trained model> -s <path that contains 3D line segments> --baseline 2

• The format of the input 3D line segments can be: .obj, .stl,
• THe format of the output 3D line segments is .stl

Command line arguments for line3d.py:

• --baseline: baseline method, default is 2, other options are 1 and 3:
  • 1: use the dataset for training the 3D Gaussian splatting model to generate trick colmap data folder, which contains images.txt, points3D.txt, and cameras.txt if the original data is in nerf_synthetic format.
  • 3: implement LineGS in different scenes, with different values of parameters in LineGS, and plot the results in figure.
• -s:
  • if the baseline is 1, the path to the dataset for training the 3D Gaussian splatting model
  • if the baseline is 2, the path to the dataset that contains the initial 3D line segments
  • if the baseline is 3, the path to the dataset for implementing LineGS, the data in the folder should contain the initial 3D line segments.

Dataset structure for baseline 3:

|- <path to dataset>
    |- <scene name 1>
        |- <colmop data folder>
            |- <3d line segments method>
                |- <name>.obj
                |- <name>.stl
            |- images.txt
            |- points3D.txt
            |- cameras.txt
    |- <scene name 2>
        |- <colmop data folder>
            |- <3d line segments method>
                |- <name>.obj
                |- <name>.stl
            |- images.txt
            |- points3D.txt
            |- cameras.txt
    |- ...

The raw data for the 3D line segments can be found in the following link: LineGS

The data contains the 3D line segments generated by Line3D++、ELSR, also the raw RGB images and corresponding colmap data folder for training the 3D Gaussian splatting model. Also contains the trained models for 3D Gaussian splatting and the results of LineGS.