TimSchneider42/abc20-sel

Twenty cylinder-family ABC objects

A hand-reviewed twenty-object subset of TimSchneider42/abc-round-50k, built as an object set for in-hand tactile classification with the Allegro hand (task.env.object.type=tactilemnist://<this repo>/train).

The point of the set is that it carries no shortcut cue. Every object is from one family -- a right circular cylinder fits each of them well -- so they differ essentially only in proportion. There is no spiky one, no flat one and no odd one out, and because the simulator rescales every object onto one shared diameter ladder, physical size says nothing about identity either.

How it was selected

From the 50 000-mesh train split, in order:

1. Cylindricity. Every mesh scored on how well a right circular cylinder fits it: for each principal axis, the circularity of the convex cross-section times the mesh volume over that cross-section extruded along the axis, best of three axes. It reads 1.000 for a cylinder, 0.785 for a box, 0.663 for a sphere, 0.332 for a cone. The pool is walked in descending order of this score.
2. Usable geometry. One watertight body of a single shell, non-degenerate, at most 60 000 faces.
3. Thickness the hand can hold. The ladder pins the caliper diameter sqrt(h^2 + d^2) to 7-11 cm, so the thinnest dimension at the worst rung is 7 * min(1, r) / sqrt(1 + r^2) cm for r = h/d. Requiring at least 3.5 cm gives r in [1/sqrt(3), sqrt(3)]; it excludes thin discs, which are plates the fingers cannot keep hold of.
4. No duplicates in proportion. For a right circular cylinder h/d is the complete rotation- and scale-invariant descriptor: two cylinders with the same h/d are the same object however the mesh is stored and whatever the ladder scales it to. No two objects are within 0.03 of each other in h/d.
5. No duplicates in appearance. Candidates are compared by the silhouette profile they present standing on their cylinder axis -- radius as a function of height, normalized by the caliper diameter, rotation about the axis factored out -- and by the same profile of their convex hull.
6. No internal holes. volume / hull_volume >= 0.97 and genus 0. A through-bore is something the hand cannot feel in this setting, so two objects differing only by a hole are the same object to the agent, and the hole is wasted shape information. Convexity alone is not sufficient (a narrow bore costs little volume), so genus is checked outright.

Steps 3-6 were added over three rounds of visual review of the rendered set; six objects were replaced in the process.

The objects

Row order is class index order, sorted by h/d.

#	id	h/d	cylindricity	volume/hull	caliper diameter (cm)
1	364954	0.584	0.772	0.854	11.91
2	189512	0.623	0.979	1.000	1.66
3	838299	0.661	0.724	1.000	18.21
4	449187	0.704	0.829	1.000	4.85
5	681312	0.759	0.725	0.770	7.24
6	347659	0.792	1.000	1.000	4.09
7	156987	0.826	0.919	1.000	17.55
8	636991	0.883	0.760	0.859	1.98
9	557510	0.917	0.712	1.000	7.03
10	229287	0.961	0.819	1.000	21.61
11	91258	1.000	0.974	1.000	6.66
12	34613	1.050	0.912	0.912	13.16
13	850197	1.151	0.879	1.000	7.13
14	341445	1.246	0.885	1.000	8.26
15	390981	1.281	0.745	0.996	18.38
16	997539	1.323	0.880	1.000	3.62
17	351236	1.426	0.966	1.000	2.86
18	535604	1.500	0.777	1.000	7.62
19	554366	1.548	0.938	1.000	16.02
20	237049	1.688	1.000	1.000	10.17

train and test contain the same twenty meshes: the object's identity is the classification label, so both splits must span the same classes. What varies between episodes is the scale rung and the initial orientation.

Columns are those of the source: id, label, mesh.vertices, mesh.faces, plus the shape statistics mean_width, diameter, roundness, aspect_ratio, volume, concavity as float32. Vertices are in meters.

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.

To identify the creator of a specific model, cross-reference the model id with the metadata dataset.