Hugging Face's logo

Datasets:
TimSchneider42
/
abc-round-50k

Tasks:
Other
Modalities:
Tabular
Formats:
parquet
Languages:
English
Size:
10K - 100K
Tags:
Mesh
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
id
int32
label
class label
mesh.vertices
list
mesh.faces
list
mean_width
float32
diameter
float32
roundness
float32
aspect_ratio
float32
volume
float32
concavity
float32
362,955
00
[ [ 0.01249999925494194, 0.012500000186264515, 0.01249999925494194 ], [ 0.01249999925494194, 0.012500000186264515, -0.012500000186264515 ], [ -0.012500002980232239, 0.012500000186264515, 0.01249999925494194 ], [ -0.012500002980232239, 0.012500000186264515, -...
[ [ 0, 1, 2 ], [ 1, 3, 2 ], [ 4, 5, 6 ], [ 5, 7, 6 ], [ 7, 5, 2 ], [ 5, 0, 2 ], [ 5, 4, 0 ], [ 4, 1, 0 ], [ 6, 7, 3 ], [ 7, 2, 3 ], [ 4, 6, 1 ],...
0.0375
0.043301
0.367553
1
0.000016
1
781,906
00
[ [ -0.004975154064595699, 0.0004978392389602959, 0.009107142686843872 ], [ -0.004993784707039595, 0.0002492294297553599, 0.009107142686843872 ], [ 4.364475037998561e-20, -1.966995510869536e-22, 0.009107142686843872 ], [ -0.004999999888241291, -1.224843452737256...
[ [ 0, 1, 2 ], [ 1, 3, 2 ], [ 2, 3, 4 ], [ 5, 6, 2 ], [ 6, 7, 2 ], [ 2, 7, 0 ], [ 8, 9, 2 ], [ 9, 10, 2 ], [ 2, 10, 5 ], [ 11, 12, 2 ], [ 12, 13, ...
0.020084
0.021213
0.549744
0.75
0.000003
1
737,185
00
[[-0.0011331761488690972,-0.0014209579676389694,-0.03650717809796333],[-0.0012025964679196477,-0.001(...TRUNCATED)
[[0,1,2],[3,4,2],[5,6,2],[7,8,9],[10,11,12],[9,8,13],[14,13,8],[12,11,13],[15,13,11],[15,11,16],[11,(...TRUNCATED)
0.059175
0.074289
0.482468
0.727232
0.000104
1
830,541
00
[[0.0015642852522432804,-0.009549999609589577,-0.0025598725769668818],[0.00391074363142252,-0.009549(...TRUNCATED)
[[0,1,2],[1,3,2],[2,3,4],[3,5,4],[4,5,6],[5,7,6],[6,7,8],[7,9,8],[8,9,10],[9,11,10],[10,11,12],[11,1(...TRUNCATED)
0.02133
0.024286
0.356095
0.785218
0.000003
0.7916
303,389
00
[[-0.001959174871444702,-1.1631271744237254e-20,-0.010427592322230339],[-0.0019567394629120827,0.000(...TRUNCATED)
[[0,1,2],[1,3,2],[2,3,4],[3,5,4],[4,5,6],[5,7,6],[6,7,8],[7,9,8],[8,9,10],[9,11,10],[10,11,12],[11,1(...TRUNCATED)
0.015113
0.017879
0.495082
0.790609
0.000001
0.922257
523,402
00
[[-0.00400021206587553,0.00020242546452209353,0.0016394476406276226],[-0.004998968914151192,0.000252(...TRUNCATED)
[[0,1,2],[1,3,2],[2,3,4],[3,5,4],[4,5,6],[5,7,6],[6,7,8],[7,9,8],[8,9,10],[9,11,10],[10,11,12],[11,1(...TRUNCATED)
0.031847
0.033801
0.306778
0.78042
0.000006
0.518292
751,603
00
[[0.0015642852522432804,-0.009549999609589577,-0.0025598725769668818],[0.00391074363142252,-0.009549(...TRUNCATED)
[[0,1,2],[1,3,2],[2,3,4],[3,5,4],[4,5,6],[5,7,6],[6,7,8],[7,9,8],[8,9,10],[9,11,10],[10,11,12],[11,1(...TRUNCATED)
0.02133
0.024286
0.356095
0.785218
0.000003
0.7916
875,069
00
[[-0.007322348188608885,-0.002610580762848258,0.0013019538018852472],[-0.007319330703467131,-0.00263(...TRUNCATED)
[[0,1,2],[3,4,5],[6,7,8],[9,10,11],[12,13,14],[13,15,14],[14,15,16],[17,18,15],[18,19,15],[15,19,16](...TRUNCATED)
0.018684
0.020378
0.37648
0.749904
0.000002
0.675073
765,766
00
[[-0.07743111997842789,0.05779922008514404,0.016852788627147675],[0.07470833510160446,0.057799220085(...TRUNCATED)
[[0,1,2],[1,0,3],[0,4,3],[3,4,5],[5,6,3],[6,7,3],[3,7,8],[9,10,11],[10,12,11],[13,10,14],[2,15,0],[1(...TRUNCATED)
0.182193
0.196167
0.57546
0.729492
0.002275
1
659,407
00
[[0.009059613570570946,-0.018268654122948647,0.017140377312898636],[0.009059613570570946,-0.01884333(...TRUNCATED)
[[0,1,2],[1,3,2],[2,3,4],[4,5,2],[5,6,2],[2,6,7],[6,8,7],[7,8,9],[10,11,2],[11,12,2],[2,12,13],[13,1(...TRUNCATED)
0.060076
0.06937
0.290741
0.706391
0.000051
0.791027
End of preview. Expand in Data Studio

Roundish ABC objects

A subset of TimSchneider42/tactile-mnist-abc-dataset (triangle meshes in meters from the ABC dataset) selected for in-hand manipulation with a multi-fingered robot hand: roughly ball-like single-piece objects of any size (the meshes keep their original dimensions; rescale them to the hand when loading).

Hard criteria (all splits):

  • exactly one connected, watertight body consisting of a single closed shell (no overlapping solids) whose volume does not exceed its convex hull's
  • no size or aspect-ratio constraint

Roundness is mesh volume / volume of the sphere with the mesh's geometric diameter (1 for a sphere, ~0.37 for a cube); the splits keep the roundest meshes:

  • train: the 50000 roundest of the 462950 source-train meshes passing the hard criteria (roundness >= 0.2886; 539624 scanned; median roundness 0.368, median diameter 5.9 cm, range 1.2 - 50.6 cm)
  • train100: 100 objects of train spread over shape and size (farthest-point sampling on z-scored diameter, mean_width, roundness, aspect_ratio, ext_mid_over_max, ext_min_over_max, concavity)
  • train20: the 20 objects of train100 spread over shape and size by the same farthest-point sampling, applied to train100 (one class per object for across-scale object classification)
  • test: the 5000 roundest of the 51445 source-test meshes passing the hard criteria (roundness >= 0.3027; 59959 scanned; median roundness 0.377, median diameter 5.7 cm, range 1.3 - 42.0 cm)
  • test100: 100 objects of test spread over shape and size (farthest-point sampling on z-scored diameter, mean_width, roundness, aspect_ratio, ext_mid_over_max, ext_min_over_max, concavity)
  • test20: the 20 objects of test100 spread over shape and size by the same farthest-point sampling, applied to test100 The train and test splits are taken from the source's splits. Rows are stored in a random order (seed 0), so any prefix is a random sample; the train100, test100 subsets keep their parent's order.

Columns: the source's id, label, mesh.vertices, mesh.faces (meshes are unchanged), plus per-mesh shape statistics mean_width, diameter (max pairwise vertex distance, rotation invariant), roundness, aspect_ratio, volume (all float32, in meters / cubic meters) and concavity (mesh volume / convex-hull volume).

License

Note that I am not the creator of the ABC dataset; hence, I hold no copyright over it.

The copyright of the CAD models is owned by their creators. For licensing details, see Onshape Terms of Use 1.g.ii.

This dataset does not contain any metadata. Hence, to identify the creator of a specific model, please cross-reference the model ID with the indices of the metadata dataset.

train20cyl / test20cyl

A hand-reviewed twenty-object subset for in-hand tactile classification, selected so that the set carries no shortcut cue: one family of solid (a right circular cylinder fits each of them well), differing essentially only in proportion, with no spiky, flat or odd one out. Since every object also enters the sim at every rung of one shared diameter ladder, size says nothing about identity either.

Selected from train by walking the meshes in descending order of a cylinder fit and keeping one only if it is usable geometry (one watertight shell, non-degenerate, at most 60k faces); thick enough for the hand to hold (at least 3.5 cm at the worst rung, i.e. h/d within [1/sqrt(3), sqrt(3)] given that the ladder pins the caliper diameter); at least 0.03 away in h/d from every object kept, h/d being the complete rotation- and scale-invariant descriptor of a cylinder; distinct in the silhouette it presents standing on its axis, and in that of its convex hull; and free of internal holes (volume / hull_volume >= 0.97 and genus 0), which the hand cannot feel in this setting.

Both splits hold the same twenty meshes: the object's identity is the classification label, so both must span the same classes. Rows are in h/d order, which is the class index order.
Downloads last month
713
Number of rows:
56,280
Total file size:
14.9 GB