Synthetic2Real Object Detection Challenge 2

This repository contains code and utilities for the Synthetic2Real Object Detection Challenge 2 on Kaggle. The goal is to train a model that predicts bounding box coordinates for objects in images, using only synthetic data.

Project Structure

_models.py
main.ipynb
view.ipynb
checkpoint.pt
sample_submission.csv
training_progress.png
train/
val/
testImages/
annotated_real_data/

• _models.py: Core model, dataset, training, evaluation, and visualization code.

• main.ipynb: Notebook for training and evaluating the model.

• view.ipynb: Notebook for visualizing predictions on images.

• checkpoint.pt: Saved model weights.

• sample_submission.csv: Example submission file for Kaggle.

• training_progress.png: Training/validation loss plot.

• train/, val/: Each contains images (.jpg or .png) and corresponding label files (e.g., .txt or .csv) for each image.
  Example:

  train/
    base/
        images/
            img_001.jpg
            img_002.jpg
            ...
        labels/
            img_001.txt
            img_002.txt
            ...
    cameraDistance/
        images/
            img_001.jpg
            img_002.jpg
            ...
        labels/
            img_001.txt
            img_002.txt
            ...
    ...
  

  The val/ and testImages/ folder follow the same structure as train/base/.

• (Optional) annotated_real_data/: Additional real annotated data.

Setup

1. Install dependencies
   This project requires Python 3.12+ and the following packages:

   • torch
   • torchvision
   • matplotlib
   • numpy
   • tqdm
   • pillow

   Install with pip:

   pip install torch torchvision matplotlib numpy tqdm pillow

2. Download and organize data
   Download the competition data from Kaggle and extract it into the appropriate folders (train/, val/, testImages/, etc.) as per the competition instructions.

Training

To train the model, open and run the cells in main.ipynb. This will:

• Load the datasets
• Initialize the model
• Train with early stopping
• Save the best model to checkpoint.pt
• Plot training and validation loss

Evaluation

After training, the notebook evaluates the model on the test set and prints the final loss.

Visualization

To visualize predictions, use view.ipynb. This notebook:

• Loads a trained model
• Displays images with predicted and (optionally) ground truth bounding boxes

Model

The model is a simple convolutional neural network defined in CoordinateCNN. It predicts the normalized center coordinates and width/height of the bounding box for each image.

License

This project is licensed under the GNU GPL v3. See LICENSE for details.

---

Contact:
For questions or issues, please open an issue on this repository.