NeuS+HashEncoding Not so Good on DTU24

[yanglixiao1994]

Hi, Benny. Have you faced the ghost floater problem when using NeuS+HashEncoding? The rendered image is good and converged to the GT. But the mesh/normal bump/sink in some areas and many floaters are on the air. encoding_config={ "otype": "HashGrid", "n_levels": 16, "n_features_per_level": 2, "log2_hashmap_size": 19, "base_resolution": 16, "per_level_scale": 1.447269237440378, "include_xyz": True, } SDF Network is nn.Linear(encoding.n_output_dims, 65) Any idea is welcome~

[bennyguo]

Hi! Are you using the default neus-dtu.yaml in the repository? Here's what I can get for this scene: I also tried training without any regularizations (distortion loss, opaque loss ...) and got similar results.

[YZsZY]

Hello Author! I recently ran into a similar situation, I replaced the occupancy grid sampling in your code, using the original resample method insteadly (nerfacc will process the results as [N_points], but I will use [N_rays,N_samples] in my follow-up work), and the rest of the code remains basically the same. But I found that my exported mesh is also very noisy, curious what causes it. Do you have any ideas? In addition the details are also very bad~

[YZsZY]

I refer to sdfstudio, xyz added to the render-net input, while the sdf-mlp output feature dimension from 13 to 257, mesh improved a lot, but there will still be some skeleton, would like to ask you has encountered this situation? Thanks a lot ~

[liuxiaozhu01]

I met the same question, but have no idea why. Have you solve it? @YZsZY