Work in progress
Changes that are built but not merged yet, and the research issues they feed. Each entry links its pull request or issue; numbers are from the PR’s own run and will move once it merges and the site rebuilds.
Piece review follow-up
PR #16, draft. Stage: fold detection and hand corrections.
- A fold added in the piece review takes the mountain or valley style printed under it, read with the same dash classifier as traced folds. The hand pick only counts where no fold is printed. On 6 of the first 15 added folds the pick disagreed with the printed line.
- After hand edits, every piece is tidied: corners and fold ends within 0.8 mm become one point, a fold with another point on it is split there, and duplicate folds or folds along the cut are dropped.
- Brings in the full review of all 178 pieces: 46 folds added, 3 dropped, 19 corners moved.
- Clean number pairs go from 1462 to 1466 of 1575.
3D fold-up against the finished model
PR #18, draft. Stage: 3D fold check (3D fold-up).
bowser.yamlgets areference:block with the finished model’s size from its public Sketchfab metadata: 1238 triangles and 633 corners.make -f fold3d.mkwritesfold3d-baseline.csv, comparing the folded trace with those counts. Corners stay doubled wherever a glue is missing or misread, so the corner difference measures how much matching is left.- On the current trace: 1375 triangles (+137) and 862 corners (+229).
Research issues
| issue | what | state |
|---|---|---|
| #11 | benchmark corpus with ground truth | synthetic templates in Benchmark; real templates to come |
| #7 | RQ1: reconstruction against a gold set and baselines | baselines A to C and metrics F1 to F5 run in Benchmark |
| #10 | RQ4: do digital metrics predict physical assembly? | metric correlations in Benchmark; 3D closure in 3D fold-up; no physical test yet |
| #8 | RQ2: ablation of joint geometry, text and topology | not started |
| #9 | RQ3: correction effort, guided vs exhaustive review | not started; corrections are logged |
The paper 1 manuscript lives in paper/ (make paper) and is not published on this site.