Stanford's Lucy
23
Likes
4300
Downloads
1759
Views
Published on January 13, 2013
This thing was Featured on January 13, 2013
Description
This was an exercise in large model simplification. The original has 28 mio. triangles, this simplified version has been reduced to 460 000. Filesize went from 1.4 GB to 21 MB.
Instructions
Recommended file: Lucy_3M_O10_20%.stl
Update2: A print of the head at 50% scale, printed at a layer height of 0.3, looks very nice. At this scale, the statue would be 70 cm high. I'm now very motivated to build a Rostock derivative, although that would normally only allow 40 cm height.
Update: The mesh was sampled with 3M points, then an out of core Poisson reconstruction applied, followed by quadric edge collapse decimation (to 20%, hence the filename). The result is a 22 MB Mesh much better than the Lucy_simplified, at no extra size. I really need to document this all logically, this might not be the best place for it.
Taken from the Stanford 3D scanning site (http://graphics.stanford.edu/data/3Dscanrep/). The Stanford CGL asks that this model not be transfomed, morphed etc.
The original file was not manifold, making a straightforward simplification in MeshLab (see thingiverse.com/thing:40632) difficult. As a workaround, the statue was subsampled to 500K points in CloudCompare (http://www.danielgm.net/cc/). A reconstrucion of the resulting point-cloud should be possible with the poisson-reconstruction function in CloudCompare. As I am more familiar in Meshlab, I used that instead.
The workflow for a point-cloud to volume conversion is as follows:
Filters > Point Set > Estimate radius from density
Filters > Point Set > Compute normals for point sets
Filters > Remeshing, etc > Surface Reconstruction: Poisson (I used Octree Depth: 9)
Netfabb repair, due to some small errors
Update: I've uploaded the script for a surface reconstruction in MeshLab (.mlx):
Filters > Show current filter script
Open script > "Surface reconstruction.mlx"
The parameters for all stages in the script can be changed beforehand, then everything can be executed at once. MeshLab also installs a server, so the script can be executed from the command line. This however is more useful when working on a lot of models at once.
Slicing at 25% scale (200 mm wide at 45°, 400 mm high) for a maximum Rostock build goes to 6 GB in RAM on my machine, so it is probably impractical to use a model with less simplification. The original model, converted to .stl, can surpass 14 GB in RAM when manipulated in Meshlab, so the point-cloud has also been included (.asc), for those wanting to try out this method. The original Model can be downloaded from the Stanford CGL site.
Update2: A print of the head at 50% scale, printed at a layer height of 0.3, looks very nice. At this scale, the statue would be 70 cm high. I'm now very motivated to build a Rostock derivative, although that would normally only allow 40 cm height.
Update: The mesh was sampled with 3M points, then an out of core Poisson reconstruction applied, followed by quadric edge collapse decimation (to 20%, hence the filename). The result is a 22 MB Mesh much better than the Lucy_simplified, at no extra size. I really need to document this all logically, this might not be the best place for it.
Taken from the Stanford 3D scanning site (http://graphics.stanford.edu/data/3Dscanrep/). The Stanford CGL asks that this model not be transfomed, morphed etc.
The original file was not manifold, making a straightforward simplification in MeshLab (see thingiverse.com/thing:40632) difficult. As a workaround, the statue was subsampled to 500K points in CloudCompare (http://www.danielgm.net/cc/). A reconstrucion of the resulting point-cloud should be possible with the poisson-reconstruction function in CloudCompare. As I am more familiar in Meshlab, I used that instead.
The workflow for a point-cloud to volume conversion is as follows:
Filters > Point Set > Estimate radius from density
Filters > Point Set > Compute normals for point sets
Filters > Remeshing, etc > Surface Reconstruction: Poisson (I used Octree Depth: 9)
Netfabb repair, due to some small errors
Update: I've uploaded the script for a surface reconstruction in MeshLab (.mlx):
Filters > Show current filter script
Open script > "Surface reconstruction.mlx"
The parameters for all stages in the script can be changed beforehand, then everything can be executed at once. MeshLab also installs a server, so the script can be executed from the command line. This however is more useful when working on a lot of models at once.
Slicing at 25% scale (200 mm wide at 45°, 400 mm high) for a maximum Rostock build goes to 6 GB in RAM on my machine, so it is probably impractical to use a model with less simplification. The original model, converted to .stl, can surpass 14 GB in RAM when manipulated in Meshlab, so the point-cloud has also been included (.asc), for those wanting to try out this method. The original Model can be downloaded from the Stanford CGL site.
You must be logged in to post a comment.
License
Stanford's Lucy by julienM is licensed under the Attribution - Non-Commercial - Share Alike license.
More from julienM
More from Sculptures
How large is the original when in printable STL format? And is there any way i could get it? I'm really interested in using this model as a detail showcase. Is it slice-ready or is there work to be done? (I've never actually worked on a model before, unfortunately i'm a complete noob when it comes to cleaning up 3d forms, etc) but i do have a capable machine with 16gb of RAM. If there is any way i can get this to a printable state i would be overjoyed!
Going with the full mesh proved difficult. Even a 5 million point subsample crashed MeshLab. This doesn't matter howerver. With a Poisson reconstruction with a depth of 11 (just in case you ever want to try) I obtained a super high resolution mesh. At 500 MB this is a bit heavy though. At depth 10, the mesh still looked very good.
I applied a quadric edge collapse decimation to 20%. That is the 2nd last image you see. The last image is a comparison of the first "_simplified" to the "sampled 3 million points, Poisson with Octree depth 10, redused to 20%". Same filesize, better result.
I've not made a printable .stl at full resolution yet. When deleting non-manifold edges and vertices, as well as deleting self-intersecting faces, there are 12.000 holes in the model. It's impractical to make MeshLab close them, as they have to be selected manually, unless I missed an option somewhere. I'm rebuilding the model from the full point-cloud as well as a 5 million subsample (5x higher). I'll upload if I succeed.
The main problem I see is Slic3r though. It burns through memory. The highly simplified model needs 6 GB. The model is 1.4 m high. Maybe if I scale before loading to Slic3r, it'l get better. Can you let me know, what build envelop you are going for? I'll then scale and upload that as well.