Dataset Viewer
Auto-converted to Parquet Duplicate
height
int64
3.55k
11.1k
max
float64
1
65.5k
min
float64
-1
0
name
stringclasses
10 values
slices
int64
9.78k
19.9k
type
stringclasses
1 value
uuid
stringclasses
10 values
voxelsize
float64
7.91
7.91
width
int64
3.4k
14.2k
format
stringclasses
1 value
7,888
1
-1
Scroll1 fiber directions
14,376
vol
scroll1-fiber-directions
7.91
8,096
zarr
7,888
65,535
0
PHerc. Paris 4, 54keV 7.91um
14,376
vol
54keV_7.91um_Scroll1A
7.91
8,096
zarr
10,112
1
-1
Scroll2 fiber directions (Joseph Balmaceda)
14,428
vol
scroll2-fiber-directions-jb
7.91
11,984
zarr
10,112
65,535
0
PHerc. Paris 3, 54keV 7.91um
14,428
vol
54keV_7.91um_Scroll2A
7.91
11,984
zarr
3,550
1
-1
Scroll3 fiber directions (Joseph Balmaceda)
9,778
vol
scroll3-fiber-directions-jb
7.91
3,400
zarr
3,550
65,535
0
PHerc0332, 7.91 - zarr - s3_raw
9,778
vol
20231117143551-zarr-s3_raw
7.91
3,400
zarr
11,136
1
-1
Scroll4 fiber directions (Joseph Balmaceda)
11,174
vol
scroll4-fiber-directions-jb
7.91
14,240
zarr
11,136
65,535
0
PHerc1667, 54keV 7.91um
11,174
vol
54keV_7.91um_Scroll4
7.91
14,240
zarr
9,016
1
-1
Scroll5 fiber directions (Joseph Balmaceda)
19,900
vol
scroll5-fiber-directions-jb
7.91
9,016
zarr
9,016
65,535
0
PHerc172, 54keV 7.91um
19,900
vol
54keV_7.91um_Scroll5
7.91
9,016
zarr

full-scrolls

Fiber orientation direction fields for the Vesuvius Challenge Herculaneum scrolls, organized to match the VC3D volpkg folder structure exactly.

What this is

Each tile contains predicted fiber orientation vectors (normal, vertical, horizontal components) for 256x256x256 voxel cubes of each scroll's CT volume. These are intended as inputs to VC3D's Surface Growth fiber-direction optimization slot, to help the segmentation pipeline trace individual papyrus sheet layers more accurately. These predictions were produced by a small 3D U-Net trained by Joseph Balmaceda on the Vesuvius Challenge fiber-skeletons dataset (31 manually annotated cubes from scrolls s1, s3, and s5). The trained model weights have not been uploaded here. The code used to train the model and generate these predictions showed these performance weights on held-out data: mean |cos(angle)| = 0.807 (1.0 = perfect, 0.5 = random). This is a baseline/proof-of-concept, not a production model.

Scroll status

  • Scroll 1 (PHercParis4): scaffold only -- already read in full by the community team (2023 Grand Prize). No predictions included.
  • Scroll 2 (PHercParis3): predictions in progress. Unread. Active prize target.
  • Scroll 3 (PHerc332): predictions in progress. Unread. Active prize target.
  • Scroll 4 (PHerc1667): predictions in progress.
  • Scroll 5 (PHerc172): predictions in progress. Note: precomputed fiber direction fields already exist at dl.ash2txt.org for parts of Scroll 5.

Structure

full_scrolls/
  Scroll1/PHercParis4.volpkg/   (scaffold only)
  Scroll2/PHercParis3.volpkg/
  Scroll3/PHerc332.volpkg/
  Scroll4/PHerc1667.volpkg/
  Scroll5/PHerc172.volpkg/
new_segments/
  run_vc.sh
  docker_commands.txt

Each volpkg:

  • config.json -- VC3D required
  • volumes/{volume_id}/meta.json -- VC3D required
  • representations/direction_fields/fiber-directions.zarr/ -- predictions
  • paths/, renders/, working/, normal_grids/ -- VC3D working dirs

Tile format

Each tile folder (z{z}_y{y}_x{x}/) contains:

  • horizontal/ -- dx component zarr array (float32, up to 256^3)
  • vertical/ -- dy component zarr array
  • normal/ -- dz component zarr array
  • meta.json -- absolute scroll coordinates, voxelsize, scroll info

VC3D usage

Load the volpkg in VC3D and point Surface Growth > Zarr folder at: representations/direction_fields/fiber-directions.zarr/horizontal

If you're gonna steal my work atleast pay me if you get a prize

Downloads last month
88