search

Fictionarry / DNGaussian

[CVPR'24] DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization
https://fictionarry.github.io/DNGaussian/
Other
279 stars 20 forks source link

apply hard depth regularization to dynamic Gaussians #34

Open Beilinlin opened 4 months ago

Beilinlin commented 4 months ago

Dear author, Hi！ if I need to apply hard depth regularization to dynamic Gaussians（4D-GS）, how should I set the opacity in render_for_depth_sh? 1721745375056 I greatly appreciate and look forward to your response and assistance.

Fictionarry commented 4 months ago

Well, I intuitively think 0.95 is still a suitable value, however, I haven't tried it on dynamic situations before. BTW, as dynamic 3dgs are actually dense view situations, whether our depth regularization can still work well for such a task is a problem. Maybe you can get better inspiration from some more related works like this: https://ambientgaussian.github.io/.

Beilinlin commented 4 months ago

Thank you very much! I have another question: Why are only loss_hard and photometric regularization used in training_sh, and not soft regularization?

Fictionarry commented 4 months ago

Thank you very much! I have another question: Why are only loss_hard and photometric regularization used in training_sh, and not soft regularization?

Actually this is because we did not apply soft depth to Blender dataset, not relevant to sh (see the hyperparameters in the script, soft depth is also not enabled in train()). The main reason is that monocular depths from depth for some scenes are extremely bad. If we force the soft depth to fit them, the quality would drop significantly. Although some scenes can still benefit from soft depth regularization, we still cancel it for concise.