WoodKnots-600: RGB images of bark-covered logs with knot instance annotations

WoodKnots-600 is a dataset for segmenting knots on the surface of round, mostly bark-covered logs in RGB images. It contains 600 images of 15 logs, with a polygon annotation for every visible knot on the log in focus. The dataset is also available on Zenodo: Zenodo DOI.

Images

The images were taken between 2022 and 2024 in Slovenian forests and show logs of four species: 8 spruce, 5 beech, 1 oak and 1 larch. Ten logs are covered with bark, three are partly debarked and two are fully debarked. The images come from several cameras, at seven resolutions between 3456×2592 and 6000×4000 pixels. The 600 annotated images were drawn at random from 2,880 captured images, so the number of images per log ranges from 16 to 87.

Annotations

Annotations are in COCO format and were made in CVAT.

• Wood knots: 1,171 non-overlapping polygons, each with the annotator's confidence score: 1.0 (definite, 653 knots), 0.6 (a defect, most likely a knot, 256) or 0.2 (possibly a defect or knot, 262). 188 images (31 %) contain no knots.
• Primary log: one polygon per image outlining the log in focus.
• Log ID for every image (groups.json).
• Knot identities for one log: 125 knot annotations in 23 images are linked to 67 physical knots, for evaluating the matching of knots across images (identity_annotations.json).

Three annotators labelled the data, and the first of them then reviewed all images to bring the labels to a common standard. A forestry expert reviewed a subset of the annotations.

Contents

Path	Description
images/	600 JPEG images
annotations.json	COCO annotations (knots, primary logs, confidence scores)
groups.json	Log ID of each image
identity_annotations.json	COCO annotations of the knots with an assigned identity (one log)

Recommended use

Images of the same log show the same knots from different viewpoints. Split the data by log rather than by image, so that no log appears in both the training and the test set. Otherwise the scores will be inflated.