search

filaPro / oneformer3d

[CVPR2024] OneFormer3D: One Transformer for Unified Point Cloud Segmentation
Other
348 stars 32 forks source link

Perform Instance Segmentation on a single .ply file. #67

Closed athanzxyt closed 3 months ago

athanzxyt commented 4 months ago

Hi!

I'm a bit confused with the readme and not sure how to proceed. I'm trying to use this model for a school project, how can I adapt this repository to perform semantic instance segmentation on a single .ply file.

oneformer3d-contributor commented 4 months ago

Unfortunately we do not provide script for it. Please looks at #9, #57.

athanzxyt commented 4 months ago

I'm still confused after reading through those previous issues. Are you saying its not possible to perform inference using the model weights you've already provided?

I don't want to train or test the model extensively, I only need to perform small inference.

oneformer3d-contributor commented 4 months ago

It is possible. You can simply do smth like build_model(config).predict(dict(points=pcd)). Where pcd is an array of n_points, 6 for (x, y, z, r, g, b). Just copy the S3DISOneFormer3D model part from our s3dis config and be sure that input point cloud is normalised in the same way as in our s3dis config.

athanzxyt commented 4 months ago

Where does build_model come from, I don't see it in either this or the mmdetection repos

oneformer3d-contributor commented 4 months ago

mmengine.registry.build_functions.build_model_from_cfg