Is there a way to run CF-GaussianObject iteratively to refine the results using more images.

[pekopull]

Hi,
I'm using CF-GaussianObject on Google Colab Pro and have encountered a limitation with VRAM.

I’d like to know if there is a way to run CF-GaussianObject iteratively so I can refine my results using more images without exceeding VRAM limits.

For context, I’m working with ceramics and pottery instead of small object, which require more images to achieve high-quality reconstructions. I want to ensure I’m getting the best possible results. I think the main VRAM bottleneck comes from Dust3r. With L4 GPU, I can process up to 24 images with Dust3r. And 32 images using A100.

If iterative refinement is not possible or hard to implement, are there any parameters I can adjust on GaussianObject to improve the quality while making better use of more compute power? (A100 GPU)

I appreciate any guidance or potential workarounds! Please also let me know more about parameters I should adjust.

Thanks!

[chensjtu]

Sorry for late reply! I also found the large cost of MASt3R and unstable results when it comes to many views. But from the other side, when the data owns many views, you can try other pose predictor for this task, i.e., vggt (https://github.com/facebookresearch/vggt); visual-sfm and et.al.