Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
1.6k
8.86k
End of preview. Expand in Data Studio

PHerc. 1447: surfaces w058, w060, w062 with ink predictions

Three consecutive windings of PHerc. 1447 in the band z = 11000–14500 of the 8.64 µm scan. Each comes with its surface geometry (tifxyz), a 24-layer surface-volume render at native resolution, and two ink predictions: one from the PHerc. 1447 finetune (ink-8um-v8in-pherc1447-loo-w062 and its sibling folds), and one from the base model ink-8um-v8in, which has never seen PHerc. 1447. w058 and w060 also include the refined ink labels used for finetuning.

w062 prediction w062 layer 12
w062 base v8-in w058 prediction
w058 base v8-in w060 prediction

Contents

Winding Render (H × W px) Area Labels Finetuned prediction (held out) Base-model prediction
w058 4512 × 8184 22.2 cm² ink + certainty mask loo_w058.png: trained on w060 + w062 v8in.png
w060 4536 × 8488 22.9 cm² ink + certainty mask loo_w060.png: trained on w058 + w062 v8in.png
w062 4668 × 8864 23.8 cm² — loo_w062.png / loo_w062_16bit.tif: LOO-w062 model, trained on w058 + w060 v8in.png

Every prediction is held out. The loo_* predictions come from three finetunes sharing one recipe (v8-in epoch 6, finetuned 10 epochs on the other two windings), each excluding the winding it predicts. v8in.png comes from the v8-in base model, which was trained without any PHerc. 1447 data, run with depth-reversed layers (--reverse). All predictions use the same tiling (64 px tiles, stride 21). Predictions are model estimates, not ground truth.

Pixel-level agreement with the refined labels of the two labelled windings (inside the certainty mask, label > 127 = ink; both predictions held out):

Winding Base v8-in AUC / AP Finetuned (LOO) AUC / AP
w058 0.862 / 0.693 0.845 / 0.658
w060 0.810 / 0.677 0.803 / 0.662
w0XX/
  layers/00.tif … 23.tif     uint8 render, one layer per normal offset (−11.5 … +11.5 voxels, step 1)
  tifxyz/x.tif y.tif z.tif   float32 surface geometry in native volume voxels, 1/4 of the render resolution
  tifxyz/meta.json           tifxyz metadata (bbox, area, scale, voxel size, source volume)
  predictions/*.png          uint8 ink probability × 255, same pixel grid as layers/
  labels/inklabels.png       ink label, 255 = ink, 0 = no ink, anti-aliased edges (w058, w060)
  labels/mask.png            certainty mask: non-zero where the label is trusted (w058, w060)
assets/                      downsampled previews
SHA256SUMS                   checksums of every file

Conventions

  • Source volume: PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr from the Vesuvius Challenge open data, resolution level 0 (8.64 µm voxels).
  • Layers: NN.tif samples the volume at normal offset NN − 11.5 voxels from the surface; layers are in ascending offset order. Both released models read them reversed (23 → 0): the finetune's config sets reverse_layers: true, and for the base model pass --reverse.
  • Geometry: tifxyz/{x,y,z}.tif give, for every grid vertex, its XYZ position in level-0 voxels. The grid has scale = 0.25, so vertex (i, j) corresponds to render pixel (4i, 4j). This is the tifxyz segment layout used by VC3D and the vesuvius Python package.
  • Predictions: the same pixel grid as the layers. Tiles are 64 px at stride 21; only tiles fully inside the render coverage are predicted, which can leave a thin unpredicted rim at the edge. The 16-bit TIFF of w062 is round(p × 65535).
  • Labels: inklabels.png / 255 is the training target (soft at stroke edges). The training loader admits a 64 px tile only when every pixel of the tile is non-zero in mask.png, so unlabelled pixels inside the mask act as confirmed no-ink. Both PNGs are RGB with equal channels; read them as grayscale.

How the surfaces were made

  1. A spiral surface model was fitted to the z = 11000–14500 band with the VC3D spiral pipeline (hand-annotated umbilicus, 3D surface patches, surface-prediction tracks and surface/normal fields). Several initialisations were compared and the best was optimised for 30,000 steps.
  2. Single windings were exported, flattened with the VC3D tools and rendered into 24 native-L0 layers. Local refinement, followed by manual editing in VC3D, fixed sheet switches, excursions into air gaps and tracking of the wrong sheet face. Because the renders kept the same UV layout, ink could be re-predicted during editing.
  3. w058 and w060 are the saved states on which the ink labels were drawn. They are the exact inputs of the finetuning runs.
  4. w062 is a later saved state, re-flattened at full resolution with SLIM (symmetric Dirichlet energy, ≤ 200 iterations, relative tolerance 1e-7). The UV mesh keeps all 3,882,056 source faces and their XYZ vertices, with no flipped or collapsed triangles. The 24 layers were rendered fresh from the new UV map. Because the re-flattening changed the image coordinates, the w062 labels drawn on the earlier grid do not align with this render and are not included.

Residual segmentation errors are possible.

Loading

import numpy as np, tifffile, cv2
w = "w062"
stack = np.stack([tifffile.imread(f"{w}/layers/{i:02d}.tif") for i in range(24)], axis=-1)  # (H, W, 24)
pred  = cv2.imread(f"{w}/predictions/loo_{w}.png", cv2.IMREAD_GRAYSCALE) / 255.0
xyz   = np.stack([tifffile.imread(f"{w}/tifxyz/{c}.tif") for c in "xyz"], axis=-1)       # (H/4, W/4, 3)

Predict a winding again with the released model:

hf download YoussefMoNader/ink-8um-v8in-pherc1447-loo-w062 --local-dir loo-w062 --exclude "training/*"
python loo-w062/predict.py --layers w062/layers --output w062_pred.png

# base model: the layers must be reversed explicitly
hf download YoussefMoNader/ink-8um-v8in --local-dir v8in --exclude "training/*"
python v8in/predict.py --reverse --layers w062/layers --output w062_v8in.png

License

CC BY-NC 4.0, following the Vesuvius Challenge data license of the underlying scan.

Downloads last month
2,147

Models trained or fine-tuned on YoussefMoNader/ink-8um-pherc1447-surfaces

Collection including YoussefMoNader/ink-8um-pherc1447-surfaces