search

chensjtu / GaussianObject

GaussianObject: High-Quality 3D Object Reconstruction from Four Views with Gaussian Splatting (SIGGRAPH Asia 2024, TOG)
887 stars 54 forks source link

About dataset #7

Open hongh0 opened 8 months ago

hongh0 commented 8 months ago

Thanks for your great work. I would like to test GaussianObject with our own dataset. Can you provide a script to handle our own dataset?

luojin commented 8 months ago

Thank you for this magnificent masterpiece. Currently, the datasets fed into this project are all massive, with each dataset being approximately 2GB or more. However, according to the paper, only four images should be utilized. The original intention of this paper was to simplify the input process, yet there is a lack of corresponding scripts for the preliminary data preparation and processing. I am eagerly looking forward to the open-sourcing of these components.

GaussianObject commented 7 months ago

Sorry for the delayed response to your issue, and appreciate your patience. To handle your own data, we recommend using COLMAP for generating poses. For estimating object masks, you can refer to SegAnyGAussians or SegmentAnything. Please let us know if you need further assistance or have more questions.

StarsTesla commented 6 months ago

@GaussianObject Hi there I was wondering，if I have four images, those images are pretty sparse, and could not generate a pose estimation by COLMAP/SFM, how should I do in such situation.

1030366507 commented 4 months ago

Sorry for the delayed response to your issue, and appreciate your patience. To handle your own data, we recommend using COLMAP for generating poses. For estimating object masks, you can refer to SegAnyGAussians or SegmentAnything. Please let us know if you need further assistance or have more questions.

The article claims that a geometric shape (GS) can be generated with just four images, but in actual operation, to create a sparse point cloud, I only have photos from four perspectives. How can I use COLMAP to generate a .sparse file with such a limited number of images? Isn't this contradictory to the title of the article? The point clouds in the dataset are all generated from a multitude of photographs. I am hoping for some guidance on how to use this project with only a few images

HermasTV commented 2 months ago

in this case a solution such as ray Diffusion can be used to generate the camera position, however I'm still investigating this hopefully create an end2end pipeline which takes 4 images and generate the 3DGS.

StarsTesla commented 2 weeks ago

@HermasTV This paper really helpful, I will check out this weekend! Thanks you very much!