Inquiry of Supporting Colmap-free 3d Gaussian Splatting

[CDchenlin]

Hi Chongjie,

Thank you so much for your insightful work. May I kindly ask if there is any plan to support Colmap-free 3d Gaussian Splatting in your studio?

Sincerely,
Chenlin

[hugoycj]

In my opinion, dust3r-based initialization could be a better choice. Colmap-free 3D Gaussian splatting only supports sequential data,and it takes a significantly longer time to reconstruct compared to using COLMAP first and then applying 3D Gaussian splatting. One advantage of colmap-free over dust3r-based initialization is that it can handle a larger number of images (e.g., more than 30). However, I believe that in the near future, this issue could be addressed by employing engineering-based methods like divide-and-conquer or better pair selection strategies.

[CDchenlin]

Hi Chongjie,

Thank you so much for your prompt response. I think building 3D scen from sequential data might be important in medical field. As we have a large number of video-based data, i.e. endoscopy. Therefore, colmap-free methods may facilitate this area.

Sincerely,
Chenlin

[Bin-ze]

Hi Chongjie,

Thank you so much for your prompt response. I think building 3D scen from sequential data might be important in medical field. As we have a large number of video-based data, i.e. endoscopy. Therefore, colmap-free methods may facilitate this area.

Sincerely, Chenlin

Hello, you can refer to my implementation. https://github.com/Bin-ze/sfm-free-gaussian-splatting
Theoretically, this can handle time series of any length. However, there is still a problem of Gaussian transmission when reconstructing a large enough scene.