LucasNgHW/AAI3001-ObjectDetection

Running the Application (app.py)

Follow these steps to run the application successfully. The app will NOT run if any step is skipped or done out of order.

1. Create a New Python Virtual Environment (REQUIRED)

It is highly recommended to isolate dependencies.

Mac / Linux

python3 -m venv venv
source venv/bin/activate

Windows

python -m venv venv
venv\Scripts\activate

You must see the (venv) prefix in your terminal before continuing.

2. Install All Required Dependencies

You MUST install using the provided requirements.txt:

pip install --upgrade pip
pip install -r requirements.txt

This installs:

• PyTorch + Torchvision
• Ultralytics YOLO
• Flask
• OpenCV
• Pillow
• Gradio (if used for deployment)
• All helper libraries used by app.py and app2.py

If installation fails, ensure you have Python 3.9–3.11.

3. Run the Application

Start the Flask app using:

python app.py

You should see output similar to:

* Running on http://127.0.0.1:5000 (Press CTRL+C to quit)

5. Open the Web Interface

Once the server is running:

1. Open your browser
2. Go to:

http://127.0.0.1:5000

You should now see the full Fruit Detection interface, including:

• Upload button
• Model selection
• Object detection results
• Bounding box highlighting
• Live Video Object Detection
• Performance Metrics

6. Optional: Running the Gradio Deployment (app.py for Hugging Face)

If you are running the Gradio version instead of Flask, use:

python app.py

You will see a Gradio URL such as:

Running on http://127.0.0.1:7860

Open that link to use the hosted interface.

7. Stopping the App

To safely stop the server:

• Press CTRL + C in the terminal.

Data Collection & Annotation

The dataset consists of 1,504 fruit images across 10 fruit classes. Data was collected through a combination of self-annotation and external sources to ensure class balance and diversity.

---

1. Self-Annotated Images (1054 images)

A total of 1504 images were annotated manually using LabelImg, of which contains around 100 images per class that were filtered from our previously collected dataset used in the first half of the project on fruit classification.

---

2. External Kaggle Sources (450 images)

To increase variability and representation, 300 additional images were sourced from multiple Kaggle datasets.
Only images that matched the 10 fruit classes were included.

All imported images were:

• Renamed and reorganized into the project structure
• Re-labelled to ensure consistent class naming
• Checked for duplicates before inclusion
• Converted any non-yolo format labels into yolo format

Kaggle Sources Used

Below are the Kaggle datasets referenced (insert actual links in the placeholders):

1. Dataset Source 1
   Fruit Images for Object Detection: https://www.kaggle.com/datasets/mbkinaci/fruit-images-for-object-detection
   Notes: 300 images of apple, banana and orange were taken from this dataset

2. Dataset Source 2
   Fruit Detection Dataset: https://www.kaggle.com/datasets/lakshaytyagi01/fruit-detection
   Notes: 150 images of watermelon were taken from this dataset

Data Splitting & Class Balancing

To ensure robust model evaluation and fair representation across classes, the dataset underwent stratified train/validation splitting and class balancing:

1. Train/Validation Split

• The dataset was split into training and validation sets using multi-label iterative stratification. This method preserves each fruit class's distribution between splits, even for images containing multiple fruit types.
• Validation size: 20% of the dataset
• All images and their corresponding YOLO label files were kept together within each split.

2. Balancing Class Distribution (Undersampling & Oversampling)

After splitting:

• Class balancing was applied to the training set only (the validation set was left untouched for realistic evaluation).
• For each fruit class, we targeted equal numbers of labeled objects (bounding boxes):
  • Undersampling: If a class had more than the target number of bounding boxes, images and labels were randomly selected so no class exceeded the target.
  • Oversampling: If a class was underrepresented, images containing that class were duplicated until the class reached the target count.
• This process ensures no model bias toward majority classes and corrects for natural imbalances after splitting.

Final training data features a balanced number of labeled objects per class, maximizing fairness and reliability in model learning.