Question about inference with new data

[npmhung]

Thank you for your great work.

I saw that in your preprocess step:

    mesh = o3d.io.read_triangle_mesh(fn)
    vertices = torch.from_numpy(np.array(mesh.vertices).astype(np.float32))
    faces = torch.from_numpy(np.array(mesh.triangles).astype(np.int64))
    superpoint = segmentator.segment_mesh(vertices, faces).numpy()
    normals = vertex_normal(vertices.numpy()[:, :3], faces.numpy())

The superpoints are extracted with segment_mesh. However, in the case when I only have point clouds as input, can I replace that with segmentator.segment_point instead? Does that affect the overall performance?

[Howard-coder191]

Yes, it’s better to use meshes as input. Using point clouds instead may slightly affect the overall performance.

[NikoloveJLP]

Dear friend, I have successfully reproduced the model based on the author's approach. Moving forward, I would like to use my own dataset. Like you, my point cloud data does not have a mesh. May I ask if you have successfully reproduced the model using your own dataset? If so, could you kindly share your experience?