search

autonomousvision / gaussian-opacity-fields

Gaussian Opacity Fields: Efficient and Compact Surface Reconstruction in Unbounded Scenes
https://niujinshuchong.github.io/gaussian-opacity-fields/
Other
559 stars 26 forks source link

Dynamic Object #15

Open hlpan opened 2 months ago

hlpan commented 2 months ago

Thank you very much for sharing such a high-quality method for reconstructing geometric models. I have found that it can restore real geometric details with high precision. However, I have also noticed some defects in the geometric models in scenes containing dynamic objects, such as moving cars being reconstructed below the ground level. I suspect this is because the moving object only appears in one image, and it is hidden in a slanted hole below the ground, making it visible from a specific angle but invisible from all others. My question is, which parameters could be adjusted or what code modifications could be made to partially solve this problem, if it cannot be completely resolved.

The moving car was captured in only one photo： image moving car reconstructed into a hole (or cave): image image

pwais commented 2 months ago

actually there was just a giant pothole

niujinshuchong commented 1 month ago

Hi, dynamic object is not supported in the current implementation but i think you can get rid of it by using object masks during training such that you ignore the loss in side the masks.