README.md

Semi-Supervised Instance Segmentation using YOLOv11

Course

CSE 438 – Image processing
Final Project

Project Overview

This project implements a complete semi-supervised learning (SSL) pipeline for instance segmentation using YOLOv11. The goal is to improve segmentation performance by leveraging both labeled and unlabeled data, addressing the high annotation cost of segmentation tasks.

The project includes:

• A supervised baseline segmentation model
• Three SSL methods:
  • FixMatch
  • MixMatch
  • Mean Teacher
• Comprehensive evaluation and visualization

---

Dataset

• Dataset: Car Damage Detection Dataset (YOLO format)
• Task: Instance Segmentation
• Source (Roboflow):
  https://universe.roboflow.com/college-gxdrt/car-damage-detection-ha5mm

Dataset Split Strategy

• 80% Training
• 10% Validation
• 10% Testing

From the 80% training split:

• 20% Labeled data
• 80% Unlabeled data

This split is used consistently across all SSL methods.

---

Models and Methods

Baseline (Supervised)

• Model: YOLOv11-seg
• Training: Only labeled data (20%)
• Purpose: Reference performance for SSL comparison
• link : https://www.kaggle.com/code/sababahoquesaba/cse438-assignment01-groupb

FixMatch

• Uses weak and strong augmentations
• Generates pseudo-labels from confident predictions
• Applies consistency regularization on unlabeled data
• link : https://www.kaggle.com/code/sababahoquesaba/fix-match-yolov11s

MixMatch

• Combines labeled and unlabeled samples
• Applies multiple augmentations
• Uses label averaging and entropy minimization
• link : https://www.kaggle.com/code/sababahoquesaba/mixmatch-yolov11s

Mean Teacher

• Teacher–student framework
• Teacher model updated using Exponential Moving Average (EMA)
• Enforces prediction consistency between teacher and student
• link : https://www.kaggle.com/code/sababahoquesaba/mean-teacher-yolov11s

Evaluation Metrics

Since this is an instance segmentation task, the following metrics are used:

• Mask mAP@0.5
• Mask mAP@0.5:0.95
• IoU-based metrics
• Bounding Box mAP (for detection comparison)

All models are evaluated on the test set.

Visualizations

The repository includes:

• Input images
• Ground truth masks
• Predictions from each SSL method
• Overlay visualizations
• Training loss and metric curves