# analog-gauge-reader

Hybrid analog gauge reader: **YOLO detect + YOLO pose** → keypoints → angle → needle fraction,  
**OCR** → scale (min/max) + unit → **final reading**.

> **Note:** OCR for digits/units on mechanical dials is inherently noisy. The angle/needle math is deterministic; OCR is used only to guess min/max/unit and can be corrected interactively.

---
## Contributors
- [@SiddharthB7](https://github.com/SiddharthB7) – detection/pose pipeline, 
- [@BevanGeorge](https://github.com/BevanGeorge) – OCR, datasets, tooling

## What this does

- **Finds the gauge** in an image (YOLO detection).
- **Finds 4 keypoints** on the dial (YOLO pose): `center`, `min`, `max`, `tip`.
- **Computes the needle fraction** using angles between `min → tip` over `min → max`.
- **Uses OCR** to estimate **scale** (`min`, `max`) and **unit** (e.g., bar, MPa).
- **Outputs final reading**: `reading = min + fraction * (max - min)` and visualizes everything.

---

## Repo structure

.
├─ gauge\_reader.py          # main pipeline & CLI (single image or folder)
├─ min\_max\_unit.py          # GaugeDetector (YOLO detection + YOLO pose)
├─ gauge\_scale\_reader.py    # OCR (OpenOCR backend) + unit parsing
├─ convert\_to\_pose.py       # helper: convert detection labels to YOLO-pose format
├─ make\_dataset.py          # helper: export keypoints/readings to CSV/Excel
├─ make\_gauge\_dataset.py    # helper: remap detection labels for single-class detect
├─ data\_det.yaml            # example YOLO detect data yaml
├─ gauge-pose.yaml          # example YOLO pose data yaml
├─ LICENSE
└─ README.md

We don’t include separate Python “training scripts” for YOLO. You train with the Ultralytics CLI (commands below). Each file has comments explaining its role.

---

Install

Create a virtual env and install the base requirements:

Windows (PowerShell)

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Linux/macOS

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

OCR backend (choose one)

This project imports:

from openocr import OpenOCR

Two different packages expose that import. Install only one:

• Recommended first: Topdu’s package

  pip install openocr-python

• Alternative: MarkHoo’s package

  pip install OpenOCR

Installing both can cause module shadowing (they both provide openocr). If you accidentally installed both, uninstall one: pip uninstall OpenOCR or pip uninstall openocr-python.

---

Run (inference)

1. Put your trained weights somewhere accessible (examples):

runs/detect/gauge-detect/weights/best.pt
runs/pose/gauge-pose/weights/best.pt

2. Open gauge_reader.py and set the two paths in main():

detect_model_path = r"...\runs\detect\gauge-detect"
pose_model_path   = r"...\runs\pose\gauge-pose"

3. Launch:

python gauge_reader.py

Choose 1 for a single image or 2 for a folder. You’ll see:

• detection box
• pose keypoints
• OCR boxes (if any)
• computed angle fraction and final value

If OCR guesses the scale wrong, the script will ask you to confirm/override min/max/unit and recompute the final reading.

---

Training (Ultralytics YOLO)

We use Ultralytics for both detection and pose.

Detection (bounding box)

yolo detect train \
  model=yolo11s.pt \
  data=data_det.yaml \
  imgsz=640 epochs=50 batch=16 \
  project=runs/detect name=gauge-detect

Pose (4 keypoints = center, max, min, tip)

yolo pose train \
  model=yolo11s-pose.pt \
  data=gauge-pose.yaml \
  imgsz=640 epochs=60 batch=16 \
  project=runs/pose name=gauge-pose

Dataset note: We started from the Roboflow gauge-analog detection dataset and converted to a 4-keypoint pose dataset. The helpers convert_to_pose.py and make_gauge_dataset.py show how labels were adapted.

---

Known limitations

• OCR is imperfect. Reading tiny/curvy/blurred digits on dials is hard; results vary with font, glare, and resolution. We expose a manual override for min/max/unit to keep the final reading stable.
• Keypoint accuracy drives the final reading. If predicted center/min/max/tip are off, the angle fraction is off. Image clarity and unusual needle shapes affect results.
• Weights & data are large. We do not commit model weights or datasets. Share via a release link or drive.

---

Acknowledgments

• Dataset: Gauge Analog (v1) — Roboflow Universe by Khang Nguyen. Used for gauge detection (bounding boxes). We converted detections to 4-keypoint pose labels for our experiments. Please refer to the dataset page for its license/terms and cite accordingly.

• Models & libs:

  • Ultralytics YOLO — detection & pose
  • OCR backend: openocr-python (Topdu) or OpenOCR (MarkHoo)

---

License

This repository is released under the MIT License (see LICENSE).