Whitefly Sticky Traps — Revised Annotations

Images of yellow sticky traps with revised whitefly annotations, ready for training object detectors (PASCAL VOC and YOLO formats), together with the data split and the YOLOv8n model used in the accompanying article. This dataset is a revision of the public dataset "Raw data from Yellow Sticky Traps with insects for training of deep learning Convolutional Neural Network for object detection" (Nieuwenhuizen et al., 2019, 4TU.ResearchData, https://doi.org/10.4121/uuid:8b8ba63a-1010-4de7-a7fb-6f9e3baf128e), released under CC0 1.0. The traps were hung in commercial greenhouses. The revision was carried out in partnership with the Brazilian Agricultural Research Corporation (EMBRAPA). Many whiteflies were missing from the original annotations. Every image was inspected at full resolution, and the whitefly (WF) class was reviewed: 1,166 missing specimens were annotated, 106 boxes were adjusted and 407 were removed, bringing the whitefly annotations from 5,591 to 6,350 across the 284 images (161 images changed). The other classes (Macrolophus, Nesidiocoris) were not systematically reviewed. The <size> field of the original XML files was also corrected to the real image resolution (5184 x 3456). Contents: - images/: 284 trap images (JPEG, 5184 x 3456 px) - annotations/voc/: PASCAL VOC XML, one file per image - labels/: YOLO TXT, one file per image (classes MR, NC, WF) - annotations/whitefly_review.csv: per-image record of the whitefly review - train.txt, val.txt: data split used in the article (227 / 57 images) - data.yaml: Ultralytics YOLO dataset config - model/best.pt: YOLOv8n model (Ultralytics 8.2.82, 1312-pixel input) On the validation images, the model reaches whitefly precision 0.865, recall 0.834 and AP@0.50 0.869, with a counting error (MAE) of 3.47 whiteflies per image. License: images CC0 1.0 (as in the original dataset); revised annotations, split and documentation CC BY 4.0. The model weights were produced with Ultralytics YOLOv8 (AGPL-3.0). See README.md inside the archive for details and citation.