About the mesh alignment and mesh evaluation

[axiub]

Thanks for the great work! We’d like to run some mesh evaluation on your dataset.
We noticed there’s an align_pisr2gt file — could you share how the alignment matrix in that file were generated?
Also, would it be possible for you to share the mesh evaluation code you used for the Chamfer distance and F-score in Table 2 of the paper?

[GCChen97]

Hi @axiub , the alignment of meshes were done by:

1. coarse alignment in blender
2. accurate alignment with ICP.

As for the metrics, I use compare_meshes to calculate CD and FS. Below is the code clip:

import open3d as o3d
from pytorch3d.structures import Meshes

def compare_meshes(...):
    # https://github.com/facebookresearch/meshrcnn/blob/main/meshrcnn/utils/metrics.py
    ...

list_th_dist = [ 0.001, 0.0005 ]

def mesh_o3d2pytorch(mesh_o3d):
        vertices = torch.as_tensor(np.asarray(mesh_o3d.vertices), dtype=torch.float32).to('cpu')
        faces = torch.as_tensor(np.asarray(mesh_o3d.triangles), dtype=torch.long).to('cpu')
        normals = torch.as_tensor(np.asarray(mesh_o3d.compute_vertex_normals().vertex_normals), dtype=torch.float32).to('cpu')
        mesh_pytorch = Meshes(verts=[vertices], faces=[faces], verts_normals=[normals])

mesh_predict_pytorch = mesh_o3d2pytorch(mesh_predict).cuda()
mesh_gt_pytorch = mesh_o3d2pytorch(mesh_gt).cuda()

mesh_predict = o3d.io.read_triangle_mesh('/path/to/estimated_mesh')
mesh_gt = o3d.io.read_triangle_mesh('/path/to/gt_mesh')

results = compare_meshes(mesh_predict_pytorch, mesh_gt_pytorch,
            num_samples='verts', scale=1., thresholds=list_th_dist, reduce=False)