DTU Mesh for Evaluation table

[mwang625]

Hi authors,

Thanks for your amazing work and quick replies to issues. I have a question about the Table 5 in the paper, about chamfer distance evaluation for DTU:

I'm wondering are the reconstructed meshes from MonoSDF in this table the same as those in the download script (https://s3.eu-central-1.amazonaws.com/avg-projects/monosdf/meshes.tar)? Because I used the meshes from dtu_allviews_mlp folder and ran eval.py under dtu_eval folder, but chamfer scores (last column in DTU.csv file, shown in the image below) cannot match with the 2nd last row in the table above.

Does it has something to do with the culling step? Below is the culled mesh for scan 65, it seems like some background mesh is not cleaned.

I'd really appreciate it if you can shed some light into this matter, thanks in advance!

[niujinshuchong]

@mwang625 Thanks for pointing this out. I think I put a checkpoint of scan37 that's trained for only 200 epochs (you can confirm by torch.load('scan37.pth')['epoch']). For all the other scenes, the quantitative results are similar which might be related to some training variants. You can download the meshes trained with 2K epochs here: https://drive.google.com/file/d/1Ty6Nu_sEHbGq6ZYALsM-oQ8Y1O_Pe22v/view?usp=share_link

[mwang625]

Thank you so much for checking and uploading the mesh!