Segmentation on my own dataset

[goyallon]

Hi @lzhengning, Thanks for releasing this code.

I would like to test your pretrained segmentation model on .obj meshes without groundtruth, do you think it is easily achievable ? Is there a particular tool you use to label the meshes and get the .json files present in your repository ?

Kind regards.

[lzhengning]

Hi, @goyallon,

To make predictions on other datasets, first remesh them using one of the tools.

The .json files is not necessary because it just helps to project the segmentation results back to the input meshes (before remeshed). If segmentation on the input meshes is also what you want, you could also write some code to do the projection and modify the test script and dataloader.

When I prepared the datasets, I used trimesh.proximity.closest_point to find the nearest triangle in the input of every face center in the remeshed. Hope this would help you.

[liang3588]

Sorry, but the "trimesh.proximity.closest_point" you mentioned cannot be opened, could you please provide the original Human Body, COSEG datasets with labeled on faces? Thank you very much!!!

[lzhengning]

@liang3588 , I have updated the hyperlink to the Trimesh document.

The original datasets can be found from their websites. Both of the them have labeled faces.

• HumanBody
• COSEG

[liang3588]

Thank you very much, but could you please provide the Human Body, COSEG datasets with 500 faces and the labels on faces?

[lzhengning]

@liang3588,

Sorry that I am a little bit confused. Do you mean the datasets provided by MeshCNN?

[unw9527]

Hi @lzhengning , I wonder how to create the .json file if I test on my own data (but maybe only one .obj file), since I want to visualize the result (see different colors produced by segmentation) and also see the accuracy. Can this be achieved easily?

If this is too troublesome, any ideas how I can simply skip the .json file and still get the colorized .ply file? Sorry that I am not quite sure what you mean in your previous comments. Thank you.

[lzhengning]

hi @unw9527 ，it won't be tricky to segment on your data.

The .json file stores face labels of the original shape and the remeshed shape. And it is used to train SubdivNet. There are three keys in a json file,

• raw_labels: the face labels in the original shape, name raw-mesh.
• sub_labels: the face labels in the remeshed shape, named sub-mesh.
• raw_to_sub: the mapping of sub-mesh and raw-mesh. The network predicts the segmentation on the sub-mesh. Sometimes, we also want to know the segmentation on the raw-mesh. The mapping is established by nearest-face query. I implemented it by trimesh.proximity.closest_point. In detail, find the closest triangle face in the raw-mesh to the center of each face in the sub-mesh. Then the index array of the triangle faces in the raw-mesh is 'raw_to_sub'.

If you do not want the accuracy, just do some modification to the source codes. Here are some functions that can help you,

• Do not use a dataloader, just load an obj file with load_mesh in SubdivNet/subdivnet/dataset.py. And pack the output to a dict like what the dataloader does.
• Use some snippet of save_results in SubdivNet/subdivnet/utils.py to save the colored segmentation results.