WoodKnots-600

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 Hugging Face: [HF link]. 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 (COCO format, 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. We recommend splitting the data by log rather than by image, because images of the same log show the same knots (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: images_[0-5].zip: 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)