YOLO Project — Data Conversion, Training, and Visualization

This repository contains utilities to convert labeled images to a YOLO dataset, train a YOLO model using the Ultralytics API, and visualize inference results in a Streamlit dashboard.

Contents

• data_convert.py — convert labelme-style JSONs into a YOLO-format dataset under YOLO_Dataset/.
• train.py — simple training entry using ultralytics.YOLO (adjust model path and training hyperparameters inside if needed).
• app.py — Streamlit app for visualizing video/image inference using trained .pt weights.
• YOLO_Dataset/ — default output folder for converted datasets (data.yaml, images/, labels/).

Requirements

Install dependencies from requirements.txt:

pip install -r requirements.txt

Make sure you have a compatible GPU driver and PyTorch/cuDNN if you plan to train with GPU.

1) Data conversion

Convert a folder tree of labeled images (LabelMe JSONs next to images) into the YOLO dataset layout.

Example:

python data_convert.py --input ./my_images_root --output ./YOLO_Dataset --ratio 0.8

What it does:

• Scans top-level subdirectories under the --input path (ignores the output folder).
• Finds image files (.jpg, .png, .jpeg) and their corresponding .json LabelMe files.
• Converts rectangle annotations to YOLO normalized bbox format and writes .txt labels.
• Copies images and labels into YOLO_Dataset/images/{train,val} and YOLO_Dataset/labels/{train,val}.
• Writes YOLO_Dataset/data.yaml with train, val, and names mapping.

Notes:

• If a subfolder has many images, the script shuffles and splits them according to --ratio.
• The converter removes the output folder if it already exists — back up data if needed.

2) Training

Training is provided by train.py using the Ultralytics YOLO API. By default it loads yolo12n.pt.

Quick start:

1. Ensure YOLO_Dataset/data.yaml exists (created by the converter).
2. Edit train.py to change the base model or hyperparameters if required.
3. Run:

python train.py

Important:

• train.py uses data=os.path.abspath('./YOLO_Dataset/data.yaml') — ensure that path is correct.
• Tune batch, imgsz, epochs, and device in train.py for your hardware.
• Training outputs are saved under runs/detect/<project>/<name> by default.

3) Visualization (Streamlit)

The Streamlit app in app.py provides two tabs: live video detection (upload video) and single image detection.

Run the app:

streamlit run app.py

How to use:

• Place trained .pt weight files under runs/detect/**/weights/ (the app searches runs/detect/**/weights/*.pt).
• Use the sidebar to select a weights file, set confidence threshold, and render width.
• Upload a surveillance video (mp4/avi/mov) for looped inference or upload an image for single-frame detection.

Notes:

• The app encodes frames as JPEG base64 for smoother browser rendering and includes simple frame-rate throttling to avoid freezing.
• If no .pt files are found, the sidebar will display an error and the app will stop.

Tips and Troubleshooting

• If data_convert.py fails to find JSONs, ensure JSON filenames match image names (same basename).
• For low GPU memory, reduce batch in train.py to 8 or 4.
• Check runs/ after training to locate saved weights (runs/detect/<project>/<exp>/weights/best.pt).

License & Contributions

This project uses permissive personal-use conventions. Feel free to open issues or suggest improvements.