Phalaenopsis Orchid Bud Images with Segmentation Labels and Manual Branch-Path Annotations
Description
This dataset provides annotated RGB images of Phalaenopsis orchid buds for instance segmentation, plant-structure analysis, and image-based growth-axis estimation. It comprises two non-overlapping image collections acquired from a total of 112 physical specimens. The primary collection contains 3,104 original multi-view RGB images acquired from 62 physical specimens, together with image-to-specimen metadata. Before augmentation, all views of each specimen were assigned exclusively to one partition, resulting in 2,328 training images, 466 validation images, and 310 test images. The images were annotated with polygon masks for the bud_leaf and darken classes and exported in YOLO segmentation format. Data augmentation was applied only to the training partition, increasing it to 11,640 images, while the validation and test partitions remained unaugmented. The final training-ready segmentation dataset therefore contains 12,416 images. The separate growth-axis collection contains 180 RGB images acquired from 50 additional specimens not represented in the primary collection. It includes 60 images from 20 single-branch specimens captured from three viewpoints per specimen and 120 images from 30 multi-branch specimens captured from four viewpoints per specimen. Manual annotations are stored in JSON and CSV formats using original-image pixel coordinates. The annotation structure supports direct Basal–Terminal (BT) paths and Basal–Junction–Terminal (BJT) paths, with multiple terminal paths retained where branching is visible. The final annotation set contains 324 JSON axis objects and 454 CSV terminal-path records. The repository also includes trained segmentation-model artifacts and a Python/Tkinter application for creating, viewing, editing, and restoring manual growth-axis annotations. The data can be reused for instance segmentation, skeleton and centerline extraction, branch-structure analysis, two-dimensional growth-axis estimation, multi-view consistency assessment, and computer-vision research for agricultural robotic perception. The manual growth-axis annotations represent two-dimensional image-space references and should not be interpreted as three-dimensional ground truth.
Files
Steps to reproduce
1. Specimen preparation and image acquisition: Juvenile Phalaenopsis specimens produced through in vitro tissue culture were removed from propagation flasks and placed on a white laboratory tabletop. RGB images were acquired with an Intel RealSense D435 camera and a UGREEN 2K webcam at approximately 30–50 cm using automatic camera settings. Specimens were manually rotated for multiple views. Images were stored as .jpg files. 2. Original RGB dataset and metadata: The dataset contains 3,104 images from 62 physical specimens. Fields in specimen_metadata.csv link each image_id to its specimen_id and view_index. Images sharing a specimen_id are different views of one specimen and are not independent observations. 3. Instance-segmentation annotation and processing: Visible targets were manually delineated in Roboflow as polygons for bud_leaf (class 0) and darken (class 1). bud_leaf denotes one distinguishable developing shoot branch and its associated leaves; darken denotes a visibly dark basal region. Only visible contours were annotated. Before augmentation, all views of each specimen were assigned to one partition: 2,328 training, 466 validation, and 310 test images. Preprocessing comprised auto-orientation and Fit resizing to 1,024 × 1,024 pixels with white-edge padding. Training-only augmentation generated five outputs per image using horizontal/vertical flips; 90° clockwise, counter-clockwise, and upside-down rotations; random rotation from −12° to +12°; shear up to ±15° horizontally and ±14° vertically; saturation ±26%; brightness ±6%; blur up to 0.2 pixels; and noise up to 0.02%. Training increased to 11,640 images; validation and test remained unaugmented. Labels were exported in YOLO segmentation format. Roboflow renamed exported files, and no manifest linking them to exact original image_id values was retained. 4. Growth-axis annotation: A separate cohort comprised 20 single-branch specimens with three views each and 30 multi-branch specimens with four views each, producing 180 images from 50 specimens. Axes were manually defined with manual_growth_axis_annotator.py in original-image pixels. B marks the proximal branch centerline, J the visible leaf-separation region, and T a visible shoot or leaf tip. Hidden structures were not reconstructed; independently emerging branches were stored as separate axes. BT and BJT annotations were saved in JSON and exported to CSV with one row per terminal path. The exports contain 324 JSON axis objects and 454 CSV terminal-path records. 5. Model reproduction and verification: Model artifacts were generated with YOLOv8s-seg in ultralytics==8.3.59 using imgsz=1024, 380 epochs, AdamW, lr0=0.001, and seed=0; exact settings are in args.yaml. Checks covered image-metadata and image-label correspondence, polygon masks, growth-axis coordinates, JSON-CSV consistency, and SHA-256 hashes. Definitions and limitations are in README.md and DATA_DICTIONARY.md.
Institutions
- Hanoi University of Science and TechnologyHanoi, Hanoi
Categories
Funders
- Hanoi University of Science and TechnologyHanoi, Hanoi