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 112 physical specimens. The primary collection contains 3,104 original multi-view RGB images 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 11,171 manually delineated polygon instances comprised 8,030 bud_leaf and 3,141 darken instances. Data augmentation was applied only to the training partition, increasing it to 11,640 images, while validation and test remained unaugmented. The final YOLO-format segmentation export contains 12,416 images and 44,362 polygon records, including augmented copies, comprising 31,865 bud_leaf and 12,497 darken records. The separate growth-axis collection contains 180 RGB images from 50 additional specimens not represented in the primary collection. It includes 60 images from 20 single-branch specimens captured from three viewpoints and 120 images from 30 multi-branch specimens captured from four viewpoints. Manual annotations are stored in JSON and CSV formats using original-image pixel coordinates. The annotation structure supports direct Basal–Terminal (BT) and Basal–Junction–Terminal (BJT) paths and retains multiple terminal paths when branching is visible. The final annotation set contains 324 axis objects and 454 terminal-path records. The repository also provides trained segmentation-model artifacts and a Python/Tkinter application for creating, viewing, editing, and restoring manual growth-axis annotations. The data support instance segmentation, skeleton and centerline extraction, branch-structure analysis, two-dimensional growth-axis estimation, multi-view consistency assessment, and agricultural robotic perception. The manual growth-axis annotations are two-dimensional image-space references and should not be interpreted as three-dimensional ground truth.

Steps to reproduce

1. Image acquisition: Juvenile tissue-cultured Phalaenopsis specimens were removed from propagation flasks and imaged on a white tabletop at 30–50 cm using an Intel RealSense D435 and a UGREEN 2K webcam with automatic settings. Specimens were manually rotated for multiple views, and images were stored as .jpg files. 2. Original data and metadata: The primary collection contains 3,104 images from 62 specimens. specimen_metadata.csv maps image_id and file_name to specimen_id, view_index, device, and dimensions. Repeated specimen_id values denote views of the same specimen, not independent observations. 3. Segmentation preparation: Targets were manually delineated in Roboflow as polygons for bud_leaf (class 0) and darken (class 1). bud_leaf denotes one developing shoot branch and its associated leaves; darken denotes a visibly dark basal region. Only visible contours were annotated. All views of each specimen were assigned exclusively to one partition: 2,328 training, 466 validation, and 310 test images. The original annotations contained 11,171 polygons: 8,309 training, 1,629 validation, and 1,233 test instances. Preprocessing comprised automatic orientation correction, contrast adjustment, reflected-edge padding, and aspect-ratio-preserving resizing to 1,024 × 1,024 pixels. Training-only augmentation produced five versions per original training image, comprising the preprocessed image and augmented variants, using horizontal/vertical flips; 90° clockwise/counterclockwise and 180° rotations; random rotation from −12° to +12°; shear from −15° to +15° horizontally and −14° to +14° vertically; saturation from −26% to +26%; brightness from −6% to +6%; Gaussian blur up to 0.2 pixels; and noise up to 0.02%. The final export contains 11,640 training, 466 validation, and 310 test images and 44,362 polygon records. raw_split.csv and roboflow_split.csv map every exported file to its source image, specimen, and partition. 4. Growth-axis annotation: A separate smartphone-acquired cohort contained 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 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 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, yielding 324 axis objects and 454 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 metadata, image-label correspondence, split provenance, 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 Technology

  Hanoi, Hanoi

Categories

Funders

• Hanoi University of Science and Technology

  Hanoi, Hanoi

Licence