Vision4Coil

A vision-based system for detecting wire rod coil tails in steel manufacturing using FFT-based temporal gating and instance segmentation. The system identifies informative coil-transition frames via frequency analysis and runs one of three interchangeable deep learning detectors to localise the coil tail.

---

Repository Structure

Top-level files

File	Description
FFT_RTSP.py	Main processing pipeline — FFT gating + detector integration + sink abstraction
detectors.py	Swappable detector classes: YOLODetector, Detectron2Detector, Mask2FormerDetector
run_on_frames.py	Batch inference script — runs any detector on a folder of validation frames
webserver.py	Flask web server — streams ROI video + FFT graph to a browser dashboard
requirements.txt	Python dependencies with install notes for Detectron2 and Mask2Former

Folders

Folder	Contents
model_weights/yolov11/	YOLO11 detection weights (best_latest.pt)
model_weights/detectron2/	Detectron2 Mask R-CNN weights (model_final.pth) + config.yaml
model_weights/mask2former/	Mask2Former weights (model_0134999.pth) + config.yaml
demo_inputs/	Sample input video (10-Coils.mov) and validation frames
demo_outputs/	Annotated output images and results.csv per model (generated at runtime)

---

Installation

1. Create a clean virtual environment

python3 -m venv vision_env
source vision_env/bin/activate

2. Install PyTorch (match your CUDA version)

# CUDA 12.8 example:
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

3. Install core dependencies

pip install opencv-python numpy plotly matplotlib pandas flask ultralytics

4. Install Detectron2

Detectron2 must be built from source. Use --no-build-isolation so the build can find your installed PyTorch:

pip install --no-build-isolation git+https://github.com/facebookresearch/detectron2.git

5. Install Mask2Former

Mask2Former is not pip-installable. Clone it and build the custom CUDA deformable attention ops:

git clone https://github.com/facebookresearch/Mask2Former.git mask2former_repo
cd mask2former_repo/mask2former/modeling/pixel_decoder/ops
python setup.py build install
cd ../../../../..

All scripts that use Mask2Former must be run with:

PYTHONPATH=$PYTHONPATH:mask2former_repo python <script>.py

---

Running the detectors

Batch inference on validation frames

# YOLO11
python run_on_frames.py --model yolo

# Detectron2 (Mask R-CNN)
PYTHONPATH=$PYTHONPATH:mask2former_repo python run_on_frames.py --model detectron2

# Mask2Former
PYTHONPATH=$PYTHONPATH:mask2former_repo python run_on_frames.py --model mask2former

Optional arguments:

Argument	Default	Description
--model	yolo	Which detector to use (yolo / detectron2 / mask2former)
--input	demo_inputs/Validation Dataset	Folder of input frames
--output	demo_outputs/<model>	Where to save annotated frames + results.csv
--conf	0.5	Confidence threshold
--weights	model-specific default	Override the default weights path

Annotated frames (mask overlay + bounding box + confidence) and a summary results.csv are written to the output folder.

Swapping detectors in the live pipeline

Edit the bottom of FFT_RTSP.py:

detector = YOLODetector()
# detector = Detectron2Detector()
# detector = Mask2FormerDetector()

Live video / RTSP stream

python FFT_RTSP.py

For a live RTSP camera feed, set credentials at the top of FFT_RTSP.py:

USERNAME = "your_username"
PASSWORD = "your_password"
CAMERA_IP = "192.168.1.100"

Web interface

python webserver.py
# Then open http://localhost:8000

Provides: live ROI video stream, FFT intensity graph, start/stop controls.

---

FFT Threshold

The FFT intensity threshold separates coil-motion frames from idle frames. Tune per coil type:

THRESHOLD = 4264.8  # DB16
# THRESHOLD = 3200  # R5.5
# THRESHOLD = 3900  # R8.5

---

Output format

When a segment ≥ 10 s is detected, a timestamped folder is created under output/:

2026_Jul_01-18-22-30_to_18-23-43/
├── *.txt        # Frequency intensity over time (tab-separated)
├── *.html       # Interactive Plotly graph
├── tail_detected_0.87.jpg   # Best frame with mask overlay + bounding box
└── tail_detected_0.87.json  # Detection metadata (class, confidence, bbox)

For batch inference (run_on_frames.py), output goes to demo_outputs/<model>/:

demo_outputs/yolo/
├── frame001_pred.jpg   # Annotated frame
├── frame002_pred.jpg
└── results.csv         # Per-frame detection results