tudelft-mdp/awesome-flowers

Awesome Flowers Dataset Collaboration Guide

This repository is maintained by TU Delft Robotics students to build a greenhouse flower detection dataset for MDP.

Please follow the rules below strictly so the dataset stays clean, merge-safe, and directly usable for YOLO training.

0) Quick Starter

Recommended: clone the repository with HTTPS

git clone https://huggingface.co/datasets/jihyopark/awesome-flowers

In HuggingFace, go to Settings -> Access tokens -> Create new token(write)

Save the token carefully, when you want to commit and push, you will need to type your username and token for authentication.

1) Collaboration Platform and Directory Structure

• Platform: Use Hugging Face Datasets as the single collaboration repository, with a Git-like workflow (clone, commit, push).
• Structure during labeling/collection:
  • Do not split into train/val at this stage.
  • Do not create per-person subfolders.
  • Everyone uploads files directly into:
    • images/
    • labels/

2) Closed-Loop Workflow: Annotate First, Then Split

Phase 1: Annotation by Contributor

To work more efficiently, not everyone needs to split the train/val/test set. Each contributor annotates locally (for example with Roboflow or Label Studio), then pushes generated images and .txt labels into shared images/ and labels/.

Phase 2: One-Click Split by Maintainer

After all contributions are submitted and image/label counts are fully matched, the project maintainer runs split.py to spilit train/val/test set. After this step, the dataset is ready for model training.

3) Mandatory Naming Rule

Non-padded numbers break lexical sorting and make dataset management error-prone. To avoid filename collisions in shared folders, every file name must follow:

f“group{id}_{frame_number}.jpg

• Zero padding: Index numbers must use fixed width (recommended: 3 digits).

Correct examples:

• group30_001.jpg

Wrong examples:

• zhang_1.jpg
• zhang_12.jpg

One-to-one pairing is required:

• Every image must have one label file with the exact same stem.
• Example: group30_001.jpg <-> group30_001.txt.

4) Data Collection and Pre-Labeling Steps

Follow this sequence before annotation:

1. Before collecting data, mix the tulips, bugs, and AprilTags in the test setup to increase variation in relative positions.
2. Capture source data (video and/or pictures).
3. Keep collection rate low (at most 5-10 frames per second) to avoid many near-duplicate images.
4. If you started from video, extract frames with ffmpeg.

Recommended ffmpeg command (extract JPEG frames at controlled FPS).

4) Annotation and Data Format: YOLO Only

• Required export format: YOLOv8 PyTorch TXT. Each line is a bbox annotation, which looks like: (class, cx, cy, h, w).
• Bounding box coordinates must be normalized floats in the range [0, 1].
• Never put all annotations into one file. Each image has its own .txt label file.

Class dictionary is single-source-of-truth:

• Class names and class IDs must strictly follow the repository root definition file (data.yaml).
• Current class mapping for this project:
  • 0: tulip_red
  • 1: tulip_white
  • 2: tulip_pink
  • 3: bug

Scenario-specific annotation rules:

• Tightness: Boxes should fit flower petals as closely as possible, minimizing background inside the box.
• Heavy occlusion: If a flower is too occluded to identify (for example, >80% occluded), treat it as background (do not label).
• Dense/ambiguous regions: Do not invent extra classes. Annotate only clearly identifiable instances using the four classes above.

6) Quick Checklist Before You Push

• Filenames follow your unique prefix + zero-padded index.
• Every image has a same-name .txt file.
• Label files are valid YOLO TXT (normalized coordinates).
• Class IDs match the repository dictionary exactly.
• Files are uploaded only to top-level images/ and labels/ during annotation phase.

7) Maintainer's workflow

To update the viewer on Hugging Face, each time there are any new commits, the maintainer should rebuild the viewer file:

git pull
python build_viewer_parquet.py

git add data/all.parquet
git commit -m "Rebuild viewer parquet"
git push