jakobkreft/spruce-log-bark-segmentation

Spruce Log Bark Segmentation

512×512 overlapping patches of Norway spruce (Picea abies) log bark with pixel-level segmentation masks for three classes. 681 image/mask pairs.

This is the training-ready subset of a larger dataset. The full collection (raw photos, processed full-resolution images, and patches at native, 1024, and 512 resolution) is archived on Zenodo: [DOI — to be added].

Classes

Masks are single-channel; each pixel holds the class index:

• 0 — bark — normal bark, from smooth to rough and scaly.
• 1 — knot — overgrown or cut-off branch marks, appearing as wart-like bumps with a "Chinese-moustache" grain pattern.
• 2 — defect — mechanical damage and peeled bark.

Class 0 dominates; classes 1 and 2 are rare.

Structure

Each patch has a matching mask sharing the same filename stem (e.g. name.png and name_mask.png).

When making train/val/test splits, split by source log, not by patch — patches overlap and several come from the same log, so a patch-level split leaks information and inflates metrics.

Source

Photographed on-site with the Biotechnical Faculty, Department of Forestry and Renewable Forest Resources, University of Ljubljana. Used to train a DiffInfinite texture model and a DeepLabv3 segmentation model.

• Code: https://github.com/jakobkreft/infinite-bark-thesis
• Models: https://huggingface.co/jakobkreft/diffinfinite-bark , https://huggingface.co/jakobkreft/deeplabv3-tree-log-segmentation

License & citation

Released under CC BY 4.0. If you use this dataset, please cite:

• TODO