Is this better?

[yuedajiong]

The range of dgcnn output is very large, useing tanh() to limit the range smaller, is this better? e.g tanh(old_output) * C (to C, for example1.0~10.0)

[liren2515]

Hello,

You should use this function to change the output of dgcnn for meshing.

[yuedajiong]

Thanks @liren2515

A related question: if I want to use another generation network to generation 'latent' from images, instead of pointcloud, does the latent using this output P_2 (P_2 = (1 - torch.sigmoid(P_1)) * udf_max_dist)) better than P_1? (My models have 2 network, one is VAE which is from PC to latent to UDF/Mesh, another is GEN which is from images to latent to UDF/Mesh. network-1 is using to help training latent-udf/mesh decoder).

Thanks.

with torch.no_grad():
latent_codes = encoder(pcds)

for i in progress_bar(range(bs), "Meshing"):
lat = latent_codes[i].unsqueeze(0)

def udf_func(c: Tensor) -> Tensor:
c = coords_encoder.encode(c.unsqueeze(0))
p = decoder(c, lat).squeeze(0) ### Here-1: abs(lat) is very large, maybe throunds ...
p = torch.sigmoid(p) ### Here-2: yes, range of p is limited.
p = (1 - p) * udf_max_dist
return p

[liren2515]

Thanks @liren2515

A related question: if I want to use another generation network to generation 'latent' from images, instead of pointcloud, does the latent using this output P_2 (P_2 = (1 - torch.sigmoid(P_1)) * udf_max_dist)) better than P_1? (My models have 2 network, one is VAE which is from PC to latent to UDF/Mesh, another is GEN which is from images to latent to UDF/Mesh. network-1 is using to help training latent-udf/mesh decoder).

Thanks.

with torch.no_grad(): latent_codes = encoder(pcds)

for i in progress_bar(range(bs), "Meshing"): lat = latent_codes[i].unsqueeze(0)

def udf_func(c: Tensor) -> Tensor: c = coords_encoder.encode(c.unsqueeze(0)) p = decoder(c, lat).squeeze(0) ### Here-1: abs(lat) is very large, maybe throunds ... p = torch.sigmoid(p) ### Here-2: yes, range of p is limited. p = (1 - p) * udf_max_dist return p

I think it won't have too much difference in terms of accuracy. Using sigmoid helps our model converge faster.

[yuedajiong]

Thanks @liren2515