mesh of real dataset - Githubissues

YanhaoZhang commented 1 month ago

Hi, I find part of the reconstructed mesh is missing. I guess tuning some parameters can be helpful. The following is a screenshot of the Bronze reconstruction (200k iteration). I use the default parameters which are also listed below. May I please ask for any suggestions?

train {
    learning_rate = 5e-4
    learning_rate_alpha = 0.05
    end_iter = 200000

    batch_size = 512
    validate_resolution_level = 4
    warm_up_end = 5000
    anneal_end = 50000
    use_white_bkgd = False

    save_freq = 10000
    val_freq = 1000
    val_mesh_freq = 10000
    report_freq = 100

    igr_weight = 0.1
    mask_weight = 0.0
}

model {
    nerf {
        D = 8,
        d_in = 4,
        d_in_view = 3,
        W = 256,
        multires = 10,
        multires_view = 4,
        output_ch = 4,
        skips=[4],
        use_viewdirs=True
    }

    sdf_network {
        d_out = 257
        d_in = 3
        d_hidden = 256
        n_layers = 8
        skip_in = [4]
        multires = 6
        bias = 0.5
        scale = 1.0
        geometric_init = True
        weight_norm = True
    }

    variance_network {
        init_val = 0.3
    }

    rendering_network {
        d_feature = 256
        mode = idr
        d_in = 9
        d_out = 3
        d_hidden = 256
        n_layers = 4
        weight_norm = True
        multires_view = 4
        squeeze_out = True
    }

    neus_renderer {
        n_samples = 64
        n_importance = 64
        n_outside = 32
        up_sample_steps = 4     # 1 for simple coarse-to-fine sampling
        perturb = 1.0
    }
}

JiaxiongQ commented 1 month ago

I think you could increase the resolution of the function ‘validate_mesh' to 512 and manually adopt a bounding box to filter mesh.

On Thu, Jul 25, 2024 at 2:39 PM Yanhao Zhang @.***> wrote:

Hi, I find part of the reconstructed mesh is missing. I guess tuning some parameters can be helpful. The following is a screenshot of the Bronze reconstruction (200k iteration). I use the default parameters which are also listed below. May I please ask for any suggestions? image.png (view on web) https://github.com/user-attachments/assets/48cd7045-2d47-4061-b62e-98c833fb1da6

train { learning_rate = 5e-4 learning_rate_alpha = 0.05 end_iter = 200000
batch_size = 512
validate_resolution_level = 4
warm_up_end = 5000
anneal_end = 50000
use_white_bkgd = False

save_freq = 10000
val_freq = 1000
val_mesh_freq = 10000
report_freq = 100

igr_weight = 0.1
mask_weight = 0.0
}

model { nerf { D = 8, d_in = 4, d_in_view = 3, W = 256, multires = 10, multires_view = 4, output_ch = 4, skips=[4], use_viewdirs=True }
sdf_network {
    d_out = 257
    d_in = 3
    d_hidden = 256
    n_layers = 8
    skip_in = [4]
    multires = 6
    bias = 0.5
    scale = 1.0
    geometric_init = True
    weight_norm = True
}

variance_network {
    init_val = 0.3
}

rendering_network {
    d_feature = 256
    mode = idr
    d_in = 9
    d_out = 3
    d_hidden = 256
    n_layers = 4
    weight_norm = True
    multires_view = 4
    squeeze_out = True
}

neus_renderer {
    n_samples = 64
    n_importance = 64
    n_outside = 32
    up_sample_steps = 4     # 1 for simple coarse-to-fine sampling
    perturb = 1.0
}
}

— Reply to this email directly, view it on GitHub https://github.com/JiaxiongQ/NeuS-HSR/issues/8, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJANJRD6GFQOQ4O42RRZOALZOCMTHAVCNFSM6AAAAABLN35N6GVHI2DSMVQWIX3LMV43ASLTON2WKOZSGQZDSMJUGY3TSNA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

YanhaoZhang commented 1 month ago

Thanks a lot for this information.

JiaxiongQ / NeuS-HSR

mesh of real dataset #8