multiqr-hackathon

DL model to detect the QR codes and get the data and classify it based on the contents of it

Problem Statement

This hackathon is about solving a real problem in the medical field – recognizing multiple QR codes on medicine packs.
Most medicine packs have more than one QR code, such as codes for the manufacturer, batch number, distributor, or regulator. The challenge is to detect all these QR codes from an image. For advanced participants, an additional bonus challenge is to decode the QR contents and classify them.

Dataset

Dataset

• You will be given a dataset of 200 medicine pack images with annotated QR bounding boxes.
• These will help you train and validate your models for the main detection task.
• A test set of 50 images will be used for final scoring, which also contains decoded QR values for bonus evaluation.
• Additional hidden images will also be used to validate accuracy and benchmarking.

Project structrure

multiqr-hackathon/
│
├── ProblemStatement.md
├── README.md                                # Setup & usage instructions
├── requirements.txt                         # Python dependencies
├── train.py                                 # Training script
├── infer.py                                 # Must implement inference (input=images → output=JSON)
├── evaluate.py                              # (Optional) for self-check with provided GT
│
├── data/   
│   ├── images                               # (participants don't commit dataset, only placeholder)
│   └── labels/                              # You’ll provide a small demo set
│
├── outputs/                       
│   ├── submission_detection_1.json          # Required output file (Stage 1)
│   └── submission_decoding_2.json           # Required output file (Stage 2, bonus)
│
└── src/                          
    ├── models/                              # Model weights (PyTorch models so .pt extension)
    |   ├── yolov8n.pt
    |   └── yolov8s.pt
    ├── annotated_augmented_dataset/         # Dataset is stored here loading & preprocessing
    |   ├── train
    |   |   ├── images
    |   |   └── labels
    |   ├── valid
    |   |   ├── images
    |   |   └── labels
    |   └── data.yaml
    ├── datasets/                            # Customm Dataset downloaded from the Google drive
    └── script/                              # Python Scripts to run any processes
        └── rename_dataset.py

Note: YOLOv8 model needs validation set while traning and testing so the test set is used as valid set testing can be done on external images

Run the project in this manner

Step 1: Set the virtual environment [Run the below commands]

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

Step 2: Download the dataset form the above link
Step 3: Annotate the images using roboflow to add the bouding box
Step 4: Use the Preprocessing and Augmentation option while versioning the dataset and download the annotated dataset inside annotated_augmented_dataset folder.
Step 5: Copy some images and their labels from test folder and store in data folder to get the inference of the images & rename the test folder as valid since yolov8 uses valid as testing internally
Step 6: Run the main.py by using this command python .\main.py it will run the below python scripts in the order
├── Sub-Step 1: Rename python script to get the clear and common naming format for images python .\src\script\rename_dataset.py
├── Sub-Step 2: Start the traning process by running this command python .\train.py
├── Sub-Step 3: Run the inference file python .\infer.py to get the inference
└── Sub-Step 4: Run the evaluate file to get the model performace details python .\evaluate

Results

Evaluation Results:
Mean IoU: 0.8500 (85% accuracy its union of QR area of predicted and real QR area)
Precision: 1.0000
Recall: 1.0000
F1 Score: 1.0000
True Positives: 29
False Positives: 0
False Negatives: 0

Training Metrics

Training Metrics per Epoch

Epoch	GPU_mem	box_loss	cls_loss	dfl_loss	Instances	Size	Box(P)	R	mAP50	mAP50-95
1/30	0G	1.879	2.122	1.393	6	640	0.775	0.855	0.818	0.393
5/30	0G	1.456	0.9304	1.189	6	640	0.891	0.921	0.829	0.524
10/30	0G	1.376	0.7205	1.150	20	640	0.894	0.915	0.840	0.503
15/30	0G	1.319	0.6614	1.125	8	640	0.895	0.912	0.861	0.577
20/30	0G	1.267	0.5947	1.110	14	640	0.878	0.915	0.838	0.547
25/30	0G	1.177	0.5322	1.090	9	670	0.899	0.915	0.861	0.581
30/30	0G	1.113	0.487	1.057	5	640	0.904	0.915	0.867	0.592

Validation Metrics

Class	Images	Instances	Box(P)	R	mAP50	mAP50-95
all	50	177	-	-	-	-

Notes

• The "Instances" column under training refers to the number of instances per batch.
• Validation metrics (Box(P), R, mAP50, mAP50-95) are reported for the entire validation set after each epoch.
• GPU memory usage is consistently 0G, indicating no GPU was used or detected.

Sample submission in output folder - The bounding box co-ordinates after the inference

Project is divided into 2 stages

Stage 1: Detection

• File name: submission_detection_1.json
• Task: Submit bounding boxes of detected QR codes.

Format:

[
  {
    "image_id": "img001",
    "qrs": [
      { "bbox": [50, 30, 120, 100] },
      { "bbox": [200, 80, 260, 140] }
    ]
  }
]

Stage 2 (Bonus): Decoding + Classification

• File name: submission_decoding_2.json (fixed name)
• Task: Submit bounding boxes along with decoded QR values and types.
• Eligibility: Only for teams that pass the Stage 1 threshold.

Format:

[
  {
    "image_id": "img001",
    "qrs": [
      { "bbox": [50, 30, 120, 100], "value": "B12345" },
      { "bbox": [200, 80, 260, 140], "value": "MFR56789" }
    ]
  }
]