Why is the reconstruction result not satisfactory?

[jiangyijin]

This is a portion of the dataset I used for reconstruction.

Below is the result of my reconstruction.

Sparse points

[jiangyijin]

Does anyone know the specific reasons why reconstruction results may not be good? Or what should I do to achieve better reconstruction results?

[fifi1ous]

You can try to get more pictures or get higher-quality pictures. If you have a low number of pictures, or low-quality pictures try to increase the number of iterations. Also, you may try to create the COLMAP database in different software for example in Agisoft metashape. I would recommend post-processing the images in some software to get better images. I had a similar problem when I had a very detailed object and it helped me.

[jiangyijin]

You can try to get more pictures or get higher-quality pictures. If you have a low number of pictures, or low-quality pictures try to increase the number of iterations. Also, you may try to create the COLMAP database in different software for example in Agisoft metashape. I would recommend post-processing the images in some software to get better images. I had a similar problem when I had a very detailed object and it helped me.

Thank you very much for your response. I used colmap for the 3D reconstruction. I was aiming to address the issue of image exposure, so I would like to understand how to modify gaussian-splatting to tackle this problem. My dataset consists of over 2000 images.

[jaco001]

GS literally put inside all what you put. So it is a better to select your pictures. Select sharp good quality pictures. No need for high resolutions. I got good results from res under 1600 pix. Usually I use x2 size that will be uses in the training.

[jiangyijin]

GS literally put inside all what you put. So it is a better to select your pictures. Select sharp good quality pictures. No need for high resolutions. I got good results from res under 1600 pix. Usually I use x2 size that will be uses in the training.GS 从字面上把你放的所有东西都放进去。因此，最好选择您的图片。选择清晰、高质量的图片。无需高分辨率。我从 1600 像素以下的分辨率中获得了良好的结果。通常我使用将在培训中使用的 x2 大小。 Thank you very much for your response. Currently, I am working on achieving three-dimensional reconstruction of extreme images. Therefore, I am interested in knowing if there are any areas where GS can be modified to address this issue.

[jaco001]

There is small group software that can 'edit' GS. Usually you need only 'trim' (use Blender plugin or websites like https://playcanvas.com/supersplat/editor ), because all colours and light comes from input pictures and are 'baked' to the model. This is great shortcut from normal methods, but also a little dead end. (for now)

Re-coloring and re-lighting you can do in games engine like Unity or Unreal Engine via plugins, but also this isn't GS anymore and you lose some quality. (bonus: Babylon 7 is js game engine https://www.babylonjs.com/ that use GS without plugin)

Converting to mesh is problematic for now. Look for 'Sugar' https://anttwo.github.io/sugar/ - it is promising