Hi, thank you for releasing the code and data for SparseOcc.
I am trying to reproduce the 3DScene experiments, especially the Totempole result in Table 3. I used the author-provided 3DScene OFF files:
- burghers.off
- copyroom.off
- lounge.off
- stonewall.off
- totempole.off
I can successfully train and evaluate the model, but my reproduced quantitative results and visual reconstruction are still different from the paper. I would like to ask for clarification about the exact 3DScene preprocessing and evaluation protocol.
My current preprocessing
Since the repository expects each scene directory to contain pointcloud.npz, I converted the provided totempole.off into pointcloud.npz by sampling points on the mesh surface.
My current conversion is roughly:
points, face_idx = trimesh.sample.sample_surface(mesh, dense_points)
normals = mesh.face_normals[face_idx]
Hi, thank you for releasing the code and data for SparseOcc.
I am trying to reproduce the 3DScene experiments, especially the Totempole result in Table 3. I used the author-provided 3DScene OFF files:
I can successfully train and evaluate the model, but my reproduced quantitative results and visual reconstruction are still different from the paper. I would like to ask for clarification about the exact 3DScene preprocessing and evaluation protocol.
My current preprocessing
Since the repository expects each scene directory to contain
pointcloud.npz, I converted the providedtotempole.offintopointcloud.npzby sampling points on the mesh surface.My current conversion is roughly: