Measurement System

(Mask R-CNN + Camera Calibration)

Overview

This project implements a full computer vision pipeline for detecting, segmenting, and measuring real-world dimensions of desired object, in this case envelope is selected using:

• Camera calibration (lens distortion removal)
• Mask R-CNN instance segmentation
• Pixel-to-mm conversion using a reference object
• Automated width/height measurement system

The system outputs real-world measurements (mm) from images using deep learning + classical vision.

---

Key Features

• Camera calibration using checkerboard images
• Image undistortion for geometric accuracy
• Envelope detection using Mask R-CNN (PyTorch)
• Pixel-to-mm conversion using reference ID card
• Automatic width & height measurement
• Visualization of results with annotations
• Full dataset pipeline (train/val/test split)
• Evaluation metrics and error analysis

---

Project Structure

xis_assessment/ ├── calibration/ ├── dataset/ ├── models/ ├── measurement/ ├── docs/ ├── requirements.txt └── README.md

more detailed in SETUP.md

---

Object Information

• Object: Envelope
• Dimensions: 139 mm × 90 mm
• Task: Instance Segmentation + Measurement
• Framework: PyTorch (Mask R-CNN)

---

Pipeline Overview

1. Camera Calibration

• Checkerboard-based calibration using OpenCV
• Computes:
  • Camera matrix
  • Distortion coefficients
• Removes lens distortion for accurate measurement

2. Dataset Preparation

• Images captured using smartphone camera
• Conditions:
  • Different lighting
  • Multiple angles
  • Varied backgrounds
• Annotation tool: CVAT (COCO format)
• Split:
  • Train: 70%
  • Validation: 20%
  • Test: 10%

3. Model Training

• Model: Mask R-CNN (ResNet-50-FPN)
• Framework: PyTorch
• Task: Instance segmentation of envelope

Hyperparameters:

Parameter	Value
Epochs	50
Batch Size	2
Learning Rate	0.005

---

Measurement System

Reference Object

• ID Card (85.6 mm × 54.0 mm)
• Used for pixel-to-mm conversion
• Detected using HSV color segmentation

Method

1. Detect reference card
2. Compute pixels per mm
3. Detect envelope using Mask R-CNN
4. Extract mask and contour
5. Fit rotated bounding box (minAreaRect)
6. Convert pixels → mm

---

Results

Model Performance

Metric	Value
mAP@0.5	1.00
mAP@0.5:0.95	0.9723
Mean IoU	0.957
Precision	1.00
Recall	1.00

---

Measurement Accuracy

Metric	Width	Height
MAE	3.23 mm	7.01 mm
MPE	3.59%	5.04%

---

Limitations

• Requires reference card in every image
• Sensitive to lighting and color similarity (HSV detection)
• Assumes flat object plane
• Accuracy decreases with increased camera distance

---

Installation

Clone Repository

git clone 
cd xis_assessment

Install Dependencies

pip install -r requirements.txt

Usage

1. Camera Calibration

python calibration/calibrate.py

2. Undistort Dataset

python calibration/undistort_dataset.py

3. Train Model

python models/train.py

4. Evaluate Model

python models/evaluate.py

5. Run Measurement Pipeline

python measurement/measurement.py

Output

Trained model: models/maskrcnn_final.pth Loss curves: models/loss_curves.png Evaluation metrics: models/evaluation_results.json Measurement results: measurement/outputs/

Technologies Used

• Python
• PyTorch
• Torchvision (Mask R-CNN)
• OpenCV
• NumPy
• Matplotlib
• CVAT (annotation tool)