CAD-to-CAD retrieval: how far do cheap geometric descriptors get you?

A shape-retrieval study on 1008 b-rep CAD parts. Query a part, get the most similar historic parts back. No learned model anywhere: a 64-bin D2 shape distribution, exact nearest-neighbour search in memory, and a confidence gauge on every result.

The retrieval is the small half. Most of the notebook goes on how you would know whether it works, on a corpus that ships no relevance labels at all.

Read the notebook — it is the report, top to bottom.

What came out of it

Corpus	1008 single-body b-rep parts (STEP), Fusion 360 Gallery segmentation subset
Method	D2 shape distribution, 64 bins, 2048 area-weighted surface samples, tessellation deflection 0.1
Storage	~0.5 MB numpy archive in memory, brute-force exact k-NN. No vector database, and the notebook argues why
Human evaluation	precision@5 = 0.387 over 277 pre-registered, blinded judgements — roughly two useful starting points in every five
Confidence signal	Spearman(NN distance, referee distance) rho = +0.331, p = 3.4e-27

Three findings worth more than the number:

The evaluation method was the thing that needed debugging. With no relevance labels, I built a referee from the dataset's per-face operation labels. It looked principled and it was wrong: a retriever using nothing but face count beat the real system on it, while returning visual garbage. The referee got demoted from judge to falsifier, and the quality claim moved onto label-free evidence plus hand labelling.

The referee and the human disagree, and the human wins. Adding a 12-scalar feature block (v1) beats the baseline on the referee for 78.3% of queries and is strictly more stable. On the 277 human judgements it is worse — 0.313 against 0.387, Wilcoxon p = 0.281, so the difference is not significant either. The baseline shipped. An automated metric that disagrees with the only ground truth you have is not a tiebreaker.

Reuse is asymmetric. This came out of hand-labelling and is the finding I did not go looking for. A retrieved part simpler than the query is not a starting point at all — the work saved can go negative. Symmetric similarity cannot express that, which is a structural limit of the whole approach rather than a tuning problem.

Running it

conda env create -f environment.yml
conda activate cad-retrieval
# then obtain the dataset -- see DATASET.md
jupyter lab cad_retrieval.ipynb

pythonocc-core is conda-forge only, which is why there is an environment.yml and no requirements.txt.

The dataset is not in this repository. DATASET.md explains how to obtain the 1008 parts from Autodesk and where to put them; filelist.txt names them.

File	What it is
cad_retrieval.ipynb	The study. Everything is here.
features.py	b-rep loading, D2 fingerprint, the 12-scalar block
evaluate.py	Tests 1-4, the seg referee, stability under re-sampling
gold_set.py	Gold-set protocol: build, then score
make_labeler.py	Browser labelling tool, one candidate at a time, blinded
make_*_figure.py	Regenerate the explainer figures
gold/labels.tsv	The 277 judgements

Limitations

• D2 is a global summary. It cannot separate a through-hole from a blind hole, four holes from four bosses, or a part from its mirror image.
• The v1 aspect ratios use an axis-aligned bounding box, so their rotation-invariance is a property of this corpus rather than of the descriptor.
• The corpus is simple single-body parts. A real archive of thin-walled, multi-body enclosures would stress a fingerprint that mostly sees the outer envelope.
• The gold set has one rater, and that rater is also the author. Blinding and pre-registering the protocol reduce that bias. They do not remove it.

Where it would go next

Cheapest first: D2's siblings from the same paper (A3, D3, D4) reuse the sampling and histogram machinery unchanged. Then a frozen b-rep-native encoder — UV-Net or BRepNet — through the same four-test harness and gold set. Topology the b-rep already holds and this study never reads: hole counts, face-adjacency statistics. A directional complexity guard, so a retrieval simpler than the query flags itself.

Origin and licence

This work began as a take-home exercise. The problem framing, the relevance definition, the evaluation design and every conclusion are my own, and no company's material is reproduced here.

Code is MIT (see LICENSE). Material derived from the Fusion 360 Gallery Dataset — the rendered part images in figures/ and in the notebook outputs, filelist.txt, and gold/ — is governed by Autodesk's dataset licence, non-commercial research use only. See DATASET.md before reusing any of it.

Christophe Hatterer