Hoang17z/fl_omr_dataset

FL-OMR-Dataset

Official dataset and pretrained models for the project:

Federated OMR for Automated Examination Paper Digitization and Assessment

This repository provides the datasets, pretrained models, template files, and detection model used in our experiments on Federated Learning for Optical Mark Recognition (OMR).

---

Repository Structure

FL-OMR-Dataset/
│
├── data/
│   ├── dataset1/
│   │   ├── roi/
│   │   └── sheets/
│   │
│   ├── dataset2/
│   │   ├── roi/
│   │   └── sheets/
│   │
│   ├── dataset3/
│   │   ├── roi/
│   │   └── sheets/
│   │
│   ├── dataset4/
│   │   ├── roi/
│   │   └── sheets/
│   │
│   └── dataset5/
│       ├── roi/
│       └── sheets/
│
├── modelAnswer/
│   ├── metadata/
│   ├── exam*.png
│   └── modelAnswer_exam*.png
│
├── weights/
│   ├── weights/
│   │   ├── warm_start/
│   │   └── local/
│   ├── weight_3_clients/
│   │   ├── centralized/
│   │   ├── FedAvg/
│   │   └── FedProx/
│   └── weight_5_clients/
│       ├── centralized/
│       ├── FedAvg/
│       └── FedProx/
│
├── best.pt
└── README.md

---

Dataset Description

The repository contains five datasets used in the experiments to simulate a highly heterogeneous federated network.

Dataset	Description
dataset1	Original OMR dataset adapted from the work of Afifi et al.
dataset2	Additional dataset collected by manually filling Dataset 1 forms to introduce distinct ink/stroke variations.
dataset3	Additional dataset featuring a new template (Exam 6) with different layouts and box sizes.
dataset4	Additional dataset featuring a new template (Exam 7) to introduce compound data shifts.
dataset5	Additional dataset featuring a new template (Exam 8) for rigorous scalability assessment.

Each dataset contains:

• roi/
  Cropped answer regions used for EfficientNet-B0 classification.

• sheets/
  Original examination sheets used for end-to-end evaluation.

---

Model Answer Templates

The modelAnswer/ directory contains:

• examination templates,
• answer keys,
• metadata required for ROI extraction and automatic grading.

These files are necessary to reproduce the complete OMR pipeline.

---

Pretrained Models

The weights/ directory contains all trained EfficientNet-B0 checkpoints organized by training scenario.

warm_start/

Warm-start model used to initialize the Local and Federated training processes (providing baseline geometric awareness prior to distributed training).

local/

Models trained independently on each client (Client 1 to Client 5) using their isolated local datasets.

weight_3_clients/ & weight_5_clients/

Global models obtained for the 3-client and 5-client scalability scenarios, respectively. Each folder contains:

• FedAvg/: Global model obtained using the Federated Averaging algorithm.
• FedProx/: Global model obtained using the FedProx algorithm (introduced to handle non-IID conditions).
• centralized/: Model trained using centralized learning on the merged training datasets (serving as the theoretical upper bound baseline).

---

YOLO Detection Model

best.pt

YOLO model used to detect answer bubbles before ROI classification.

---

Experimental Pipeline

The complete OMR pipeline consists of the following stages:

1. Detect answer bubbles using YOLO.
2. Align examination sheets with the template.
3. Crop answer regions (ROI).
4. Classify each ROI using EfficientNet-B0.
5. Generate the final grading results.

---

Acknowledgement

Part of dataset1 is adapted from the OMR dataset introduced by Afifi et al.

The additional datasets (dataset2 to dataset5) were collected and constructed for evaluating Federated Learning under compound non-IID heterogeneous data distributions (spanning geometric, statistical, and optical variations).