How to use with my own data

[douglas0406]

Thanks for your excellent work! I would like to see the results with my own data, which includes a set of images and corresponding camera poses in Blender format. How should I use your project? I have already set up the project environment. Thank you for your patient response. I use images generated from Tencent Hunyuan3d model, the camera parameters are: +0, +60, +120, +180, +240, +300.I want to see if this works with your method. Thanks a lot!

[tlind]

Thanks for the great work. I would be interested in the same. Would be great if you could provide an example for using a custom dataset, or provide some pointers!

Update: I saw that this comment might provide some starting points: https://github.com/donydchen/mvsplat360/issues/3#issuecomment-2481906076

[donydchen]

Hi, @douglas0406, thanks for your appreciation.

MVSplat360 is mainly designed and trained on scene-level datasets (with background, not specifically orbiting one object), and I'm not sure how well it will perform on the object-level data you provided.

For scene-level data, you can use COLMAP to get the corresponding camera poses, and related instructions can be found at https://github.com/donydchen/mvsplat360/issues/3#issuecomment-2481906076, as mentioned by @tlind above.

If COLMAP fails to get camera poses due to limited input views, you might refer to VGGSfM, which might be more robust and also provides COLMAP-like outputs. For converting COLMAP raw data to NeRF/3DGS required inputs, more instructions can be found in our previous work MatchNeRF.

Hope it helps. Cheers.

[trungphien]

Hi @donydchen , is COLMAP effectively used for scene-level data? If not, could you recommend another SFM algorithm that works well for scene-level data? Thank you very much

[donydchen]

Hi @trungphien, COLMAP is effective. You can also consider using VGGSfM.