3D Gaussian Splatting is great, but can it work without the pre-computed camera poses? Introducing: COLMAP-Free 3D Gaussian Splatting Our recent work shows not only it can, but 3D Gaussians make camera pose estimation easy (compared to NeRF) along with reconstruction. 👇🧵

Starting from the initial frame, we process the following frames in a sequential manner by progressively growing the 3D Gaussians set and estimating the camera pose at the same time. 2/n)

Our approach demonstrates a significant improvement in novel view synthesis, enhancing both quality and efficiency, compared to sota unposed NeRF. 3/n)

Additionally, we show enhanced effectiveness and robustness in camera pose estimation, especially in challenging scenarios like 360 videos. 4/n)

Check out our website: arxiv: Collaboration with @yangfu21, Sifei Liu, Amey Kulkarni, @jankautz, Alyosha Efros and thank you for @_akhaliq for sharing:

It's a truly innovative and neat approach to leverage DPT depth for camera pose estimation within the 3DGS framework. Is there a schedule for releasing the code this month? :)

In addition to looking at Dr. Wang's work, it may be worth while to check out research on SMERF that was also released today.

@PMel3D Thanks for the note. In general colmap was used for preprocessing in these applications. We want to humbly emphasize this is helping build 3D Gaussian without colmap. But we do not exclude the possibility there are better sfm approaches out there.