Buckets:
| Name | Size | Uploaded | Xet hash |
|---|---|---|---|
| labels | 380,991 items | ||
| README.md | 10.7 kB xet | 91e7a61e |
9 µm ink detection labels
Ink labels for training Herculaneum ink-detection models at ~9 µm isotropic resolution, across four scrolls. Labels only - no CT data; surface volumes come from the open-data server. The models trained on these labels live at scrollprize/ink_9um. Training and inference code: villa @ merge-ink-pipelines - see ink-detection/configs/README.md
Layout
ink_9um/
labels/
aligned-scrollprizeorg-21slices/ 24 segments (PHerc0139, PHerc1667, PHercParis4, PHerc0814)
<segment>/
<segment>_inklabels.zarr
<segment>_supervision_mask.zarr
<segment>_validation_mask.zarr (3 segments only)
native9-scrollprizeorg-21slices/ 5 segments (PHerc0139 w035 w039 w040 w041 w044)
<segment>/
<segment>_inklabels.zarr
<segment>_supervision_mask.zarr
Labels
The two folders are two different sources at the same effective scale:
aligned-scrollprizeorg-21slices: annotations transferred onto public 2.4 µm surface volumes, then pooled 4x in XYZ to ~9.6 µm. Arrays are 21 slices deep, annotated only at Z=10; the other planes are zero.native9-scrollprizeorg-21slices: annotations made directly on native 9.362 µm PHerc. 0139 volumes. Arrays are 28 slices deep, annotated only at Z=14.
They stay separate because the training configs read them as two datasets entries, each with its own segments_path. Point the entries at the two folders:
"datasets": [
{
"segments_path": "/path/to/ink_9um/labels/aligned-scrollprizeorg-21slices",
...
},
{
"segments_path": "/path/to/ink_9um/labels/native9-scrollprizeorg-21slices",
...
}
]
The training loader finds <segments_path>/<segment>/<segment>_inklabels.zarr and friends on its own, so the layout here is exactly what a training run consumes. The three _validation_mask.zarr segments (pherc0139-w016, pherc0814-46527, pherc1667-w029) are the online-validation cases the released checkpoints report metrics on
Annotations originate from the Vesuvius Challenge ink annotation dataset and were transferred onto the public surface volumes with the tifxyz_label_transfer scripts in villa (volume-cartographer/scripts/tifxyz_label_transfer), with a per-case offset review
Source volumes
The labels annotate the public z-anisotropic 2.399 µm volumes downsampled to ~9.6 µm isotropic (level-2 XY, 4x z mean-pooling), with segments from PHerc. 0139 (9), Scroll 1667 (6), PHerc. Paris 4 (8), and PHerc. 0814 (1), plus 5 native 9.362 µm segments from PHerc. 0139. See the model README for how inputs are prepared
Aligned public 2.399 µm volumes (used at level-2 XY with 4x z mean-pooling):
Native 9.362 µm volumes (PHerc. 0139, used as-is):
| Segment | Scroll | Source volume |
|---|---|---|
| w035 | PHerc. 0139 | 20260317000000-w035_2026031718 · 9.362 µm |
| w039 | PHerc. 0139 | 20260302000000-w039_2026030210 · 9.362 µm |
| w040 | PHerc. 0139 | 20250831000000-w040_2025083102 · 9.362 µm |
| w041 | PHerc. 0139 | 20260108000000-w041_2026010816 · 9.362 µm |
| w044 | PHerc. 0139 | 20260115000000-w044_2026011522 · 9.362 µm |
Note: segment names in the first column come from the annotation dataset and do not track the public segment w-numbering - e.g. pherc0139-w016 maps to the public segment 20250108000004-w029
- Total size
- 3.68 TB
- Files
- 11,657,081
- Last updated
- Aug 18
- Pre-warmed CDN
- US EU US EU