The Segmentation Dataset contains a segmentation of design bodies based on the modeling operation used to create each face, e.g. Extrude, Revolve Fillet and Chamfer.
Segmentation data is extracted from the parametric feature timeline of each Fusion 360 CAD model. We choose a small subset of the most common modeling operations: extrude, chamfer, fillet, and revolve, and record the faces that were generated by these modeling operations.
In order to create a segmentation which contains as much information as possible about the CAD modeling operations we further subdivide the feature categories as follows. The faces created by extrude and revolve operations are separated according to whether they are at the side of the extrusion or at its start or end. We also divide extrude operations into additive (i.e. adding) and subtractive (i.e. cutting) extrusion operations.
The set of eight possible labels for each face are:
ExtrudeSideExtrudeEndCutSideCutEndFilletChamferRevolveSideRevolveEnd
The segmentation dataset contains a total of 35,680 3D models in three different representations: B-Rep, mesh, and point cloud.
The breps folder contains solid models (B-Rep) in two formats. The breps/step folder contains STEP files which can be read and processed by many open source tools like Open Cascade. For an example of how to read the STEP data and generate input for a neural network see the BRepNet repository. The data processing pipeline is explained in more depth here.
The breps/smt folder contains the solids in the Autodesk Shape Manager solid text format (.smt). These files can be read into Fusion 360 and other Autodesk products. The Fusion 360 API gives extensive access to the underlying B-Rep data structure.
Each brep file is accompanied by a segmentation file (.seg) in the breps/seg folder. This is an ascii text file containing the segment indices for each face in the B-Rep data. The same segfiles can be used for both the smt and step data. The file segment_names.json in the root folder contains segment names corresponding to each segment index.
For an example of how to view the STEP data and segmentation see this notebook.
The timeline_info folder contains JSON data describing the CAD modeling features which generated each B-Rep face. This information is extracted from the parametric timeline of the original design. The order of faces in the json file will match the order of faces when accessed via the Fusion 360 API. For each face we include a unique identifier for the CAD modeling feature which created it. In addition, faces generated by extrude or revolve features are marked as either start faces, end faces or side faces.
"faces": [
{
"feature": "91e05b8a-89d0-11ea-bb1a-54bf646e7e1f",
"location_in_feature": "SideFace"
},
{
"feature": "91e05b8a-89d0-11ea-bb1a-54bf646e7e1f",
"location_in_feature": "StartFace"
},
{
"feature": "91e05b8b-89d0-11ea-a01e-54bf646e7e1f",
"location_in_feature": "EndFace"
}
]
Additional information about the CAD modeling features used to construct each model is also included in the json data. The key for each CAD modeling feature is the unique identifier used in the faces array. The data contains the feature's name, type and index in the timeline.
"features": {
"91e05b8a-89d0-11ea-bb1a-54bf646e7e1f": {
"name": "Extrude1",
"type": "ExtrudeFeature",
"operation": "NewBodyFeatureOperation",
"timeline_index": 2
},
"91e05b8b-89d0-11ea-a01e-54bf646e7e1f": {
"name": "Extrude2",
"type": "ExtrudeFeature",
"operation": "CutFeatureOperation",
"timeline_index": 3
},
"91e0829c-89d0-11ea-bc45-54bf646e7e1f": {
"name": "Chamfer1",
"type": "ChamferFeature",
"timeline_index": 6
}
}The feature type can be one of:
For extrude and revolve features we also provide the modeling operation used to combine the new geometry with the rest of the model. The operation type is defined by FeatureOperations and can be one of:
NewBodyFeatureOperationJoinFeatureOperationCutFeatureOperationIntersectFeatureOperation
The meshes folder contains high quality meshes in .obj format. The meshes are guaranteed to be watertight and manifold. The accompanying .seg text file gives a segment index for each triangle in the obj data, corresponding to the eight labels in the segment_names.json file. The .fidx text file gives the face index for each triangle in the obj data. These indices will point into the array of faces in the timeline_info/*.json files.
For an example of how to visualize the mesh data and ground truth segmentation please see this notebook.
The point_clouds folder contains point clouds with 2048 samples generated randomly with an even distribution of over the surface of the triangle mesh. Each row of the .xyz files contains the x, y, z of the point and unit normal of the triangle the point was drawn from. The .seg text file gives the segment index for each point and the .fidx file gives the face index for each point.
All models in the dataset have been translated and scaled based on the axis aligned bounding box of the geometry. The smt, meshes and point_clouds have been translated so the center of the bounding box is moved to the origin and an isotopic scaling factor applied so that the longest edge of the bounding box fits into the range [-1.0, 1.0].
Due to a unit conversion in the data exchange process, the STEP files are scaled by a factor of 10. i.e. the longest side of the bounding box for the STEP files is [-10, 10]. See here for an example of how to scale the STEP data into the range [-1.0, 1.0].
To restrict the dataset to a limited number of segmentation classes we 'suppress' some CAD modeling features, slightly modifying the design from its original state.
The official train/test split is contained in the file train_test.json. The training set contains 30,459 models with the remaining 5,399 in the test set.
The extended STEP dataset contains 42,912 STEP files with all the associated segmentation information (seg files and timeline_info). This includes all the STEP data from s2.0.0, along with some models which could not be meshed with an edge count close to 2500 and consequently were not used as part of the MeshCNN baseline in the paper BRepNet: A Topological Message Passing System for Solid Models. This graph shows how the distribution of faces and modeling operations compares in the two datasets. While the overall distributions are similar, the additional bodies added to the extended dataset help fatten the long tail of more complex models.
The extended STEP dataset is provided for the B-Rep representation only. It is recommended for use cases where comparison with the mesh and point cloud representations are not required. The layout of the folder is identical to s2.0.0, but the meshes and point_clouds folder is not included.
The file additional_breps.json contains a flat list of files which were not included in the main s2.0.0 dataset. The file additional_breps_train_test.json gives the official train/test split for these additional files. The file train_test.json gives the full train/test split for the entire extended dataset. The dataset can be used with the scripts in the BRepNet repository as a drop-in replacement for s2.0.0.
| Version | Designs | Notes |
|---|---|---|
| s1.0.0 - 3.1 GB | 35,858 | This version did not contain the data as step files. It contains an additional 178 designs which could not be converted to STEP with consistent labels. In this version of the dataset the breps folder contains both smt and seg files. |