FCP parts: mesh vs point cloud

fcp v5.2 (cuda), defaults: 1024 atoms, 128 layers, parts="tracked-fit", fit_tol="auto"; sliced upright: up = the world axis each object stands on (+z for all). Point clouds: 200,000 points sampled uniformly on the source mesh (seed 0, no noise). Errors: symmetric nearest-point distance between 50k samples of the source and of the decoded mesh, over the source's bounding-box diagonal. Generated 2026-10-27 21:56 by scripts/parts_report.py.

Studies: SYNTHESIS GWN speed thin walls final gallery full gallery analysis: A vs B middle slices: A vs B parts clean 10000 noisy slices bowl gso 10k slices loaf pan 10k bowl gso 10k refined loaf pan 10k refined bowl of 10k bowl of 10k refined loaf pan 107-115 refinement splits dino

1. Parts per encoding

"point test" (prototype, scripts/pointfit.py, package untouched): a piece of a track stays whole when, at every layer, one height fit's ring is within 2 spacings (p95) of the slicer's ring in explaining the layer's points, and the fitted ring lies on the points within the same margin (both directions). Where to cut and the minimum piece length are the package's.

Red: a point-cloud encoding with more than 1.5× the parts of the same object's mesh encoding. "short": parts spanning ≤ 3 layers. spacing and fit_tol are in canonical units, from the descriptor's provenance.

object	input	up	watertight	parts	holes	short	median span	bytes	spacing	fit_tol	encode s	mean/diag	p95/diag
battery_stepped	mesh	+z	yes	4	0	0	14	563			0.71	0.00329	0.00627
battery_stepped	points 200k	+z	yes	5	0	0	16	6212	0.4722		0.69	0.00331	0.00627
battery_stepped	points 200k, point test	+z	yes	4	0	0	15	6209	0.4722		3.39	0.00335	0.00629
dress	mesh	+z	yes	17	0	0	13	6248			0.18	0.00219	0.00417
dress	points 200k	+z	yes	23	0	3	9	6266	0.3069	0.01108	0.64	0.00235	0.00464
dress	points 200k, point test	+z	yes	18	0	3	10	6251	0.3069	0.01108	2.62	0.00234	0.00467
elephant	mesh	+z	yes	27	3	11	6	6278			0.18	0.00410	0.00839
elephant	points 200k	+z	yes	28	2	8	6	6281	0.9039		0.69	0.00414	0.00842
elephant	points 200k, point test	+z	yes	26	2	8	5	6275	0.9039		3.46	0.00407	0.00827
bowl_of	mesh	+z	yes	5	2	0	19	6212			0.06	0.00263	0.00512
bowl_of	points 200k	+z	yes	6	2	1	13	6215	0.4504		0.80	0.00263	0.00512
bowl_of	points 200k, point test	+z	yes	6	2	1	13	6215	0.4504		5.62	0.00263	0.00512
bowl_gso	mesh	+z	yes	6	2	0	10	6215			0.06	0.00262	0.00508
bowl_gso	points 200k	+z	yes	5	2	0	21	6212	0.6345		0.87	0.00263	0.00505
bowl_gso	points 200k, point test	+z	yes	5	2	0	21	6212	0.6345		6.19	0.00263	0.00505
mug	mesh	+z	yes	8	2	0	22	6221			0.06	0.00292	0.00565
mug	points 200k	+z	yes	8	2	1	22	6221	0.5588		0.76	0.00292	0.00564
mug	points 200k, point test	+z	yes	8	2	1	22	6221	0.5588		5.05	0.00292	0.00564
dino	mesh	+z	yes	25	0	5	9	6272			0.07	0.00223	0.00416
dino	points 200k	+z	yes	40	0	5	8	6317	0.4802		0.79	0.00263	0.00523
dino	points 200k, point test	+z	yes	28	0	5	8	6281	0.4802		3.50	0.00256	0.00514
shoe	mesh	+z	yes	22	5	4	9	6263			0.09	0.00299	0.00557
shoe	points 200k	+z	yes	28	5	5	7	6281	0.7319		0.86	0.00312	0.00584
shoe	points 200k, point test	+z	yes	28	7	5	6	6281	0.7319		5.17	0.00311	0.00583
loaf_pan	mesh	+z	yes	7	2	0	6	6218			0.08	0.00295	0.00551
loaf_pan	points 200k	+z	yes	8	3	0	5	6221	0.893		0.83	0.00298	0.00555
loaf_pan	points 200k, point test	+z	yes	7	2	0	5	6218	0.893		5.84	0.00299	0.00556
horse	mesh	+z	yes	18	0	4	8	6251			0.07	0.00217	0.00398
horse	points 200k	+z	yes	24	0	4	9	6269	0.4621		0.74	0.00228	0.00419
horse	points 200k, point test	+z	yes	20	0	4	8	6257	0.4621		3.58	0.00227	0.00420

2. Renders

Canonical frame, orthographic, height axis z up (layer 0 at the bottom). Left: rings of every part seen along +y (x right; one colour per part, holes dashed; tall parts thinned to ~40 rings). Then the decoded mesh along +y (x-z) and along -x (y-z), each face coloured by the part whose rings are nearest (matplotlib: expect some depth-sorting artefacts). Right: each part's span of layers (holes hatched).