CabbageRowSeg: A Cross-Row Cabbage Instance Segmentation Dataset from Repeated Field Observations

CabbageRowSeg contains 929 RGB images at 1280 by 720 pixels and 1,281 cabbage mask occurrences from repeated views of approximately 60 physical plants in three rows during one field session. Row-organized data: R1 has 344 images and 542 masks; R2 has 295 images and 383 masks; R3 has 290 images and 356 masks. Each row has separate images and labels directories. Three row-disjoint folds are defined: F1 trains on R1, validates on R2 and tests on R3; F2 trains on R2, validates on R3 and tests on R1; F3 trains on R3, validates on R1 and tests on R2. The dataset-only package provides the original PNG images, YOLO polygon labels, existing COCO row annotations, fold definitions, image lists, documentation and integrity checksums. All original image and YOLO label bytes are unchanged from the previous release. The 929 frames and 1,281 annotations are repeated observations, not independent plants. Trained model checkpoints, pretrained weights, model configurations, prediction outputs, training logs and benchmark evaluation code are not included. The dataset is licensed under CC BY 4.0.