chensjtu / GaussianObject

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

Do you guys are aware for Marigold Depth ? #13

Open al3dv2 opened 7 months ago

al3dv2 commented 7 months ago

Hi I saw that you guys use Zoe depth but Marigold Depth have higher resolution https://huggingface.co/spaces/toshas/marigold May be it's can help to get even better result ? thanks for this awesome project

GaussianObject commented 7 months ago

Thank you for bringing this to our attention! We are in the process of evaluating the Marigold Depth model within our framework. We share your optimism and hope that incorporating Marigold will indeed enhance our method's performance.

al3dv2 commented 7 months ago

@GaussianObject there's now better than marigold https://github.com/kijai/ComfyUI-depth-fm?tab=readme-ov-file

GaussianObject commented 7 months ago

Thank you again for your valuable input! It's exciting to see the community engaging and suggesting alternatives that might enhance our project. We've taken a look at the link you provided for ComfyUI-depth-fm, and we appreciate the recommendation.

To add to our previous discussion, we've conducted additional ablation studies on mip360, focusing on the impact of depth loss. Here are the experimental results:

Method LPIPS PSNR SSIM
w/o_depth_loss 5.0860 24.84 0.9341
GaussianObject 4.9926 24.81 0.9350

These results indicate that depth loss doesn't play a dominant role in our methodology. Based on this insight, we are exploring how to better integrate depth estimation models and Gaussian point clouds to further improve our system.

Your feedback has been incredibly helpful, and we look forward to possibly integrating these insights into our work. Thank you for your contribution to our project's development!

al3dv2 commented 7 months ago

Thank you! I've found another one that might be better than other depth methods, and it's do normal too https://github.com/kijai/ComfyUI-Geowizard

smoothumut commented 6 months ago

Sincerely, many thanks for this great project. We will also be looking forward to the updates regarding the revised model with the new ways of integrated depth estimation models.