Benchmark
The paper 1 benchmark (make bench, make bench-report) asks whether a vectorization that looks right is also right to cut. It runs existing raster-to-vector methods on synthetic papercraft templates with exact ground truth, at three raster qualities, and scores each result with conventional visual metrics (raster IoU, chamfer and Hausdorff distance) and the fabrication-aware metrics F1 to F5 from the paper. The T- rows are the ground truth with one known defect added (an open contour, folds drawn as cuts, a 2 % scale error, 0.4 mm jitter), so the metrics can be checked against errors of known size.
12 templates, 13 pages, 3 raster conditions, 8 methods.
Visual quality does not predict fabrication quality
Mean per method and condition
Spearman correlation between visual and fabrication metrics
