Simple working example of a convolutional neural net implemented in PyTorch that predicts the center of a circle in a black and white image.

The project is organized into the following key components:

Synthetic Data Generation: Involves generating black and white images with circles of random sizes and positions, and storing the coordinates of their centers.

CircleNet Model: A CNN model implemented in PyTorch. The model takes an image as input and outputs the coordinates of the circle's center.

Training Procedure: The model is trained on the synthetic dataset with the aim to minimize the discrepancy between the model's predicted circle center and the true circle center.

Prediction & Visualization: The prediction script loads a trained CircleNet model and performs predictions on a new set of images. Predicted circle centers and true circle centers are visualized for comparison.

git clone https://github.com/alxschwrz/circlenet-cnn-pytorch.git
cd circlenet-cnn-pytorch
pip3 install -r requirements.txt

The script "generate_synthetic_spheres.py" can be used to generate a synthetic dataset of black and white images, each containing a circle of random size and position. Along with each generated image, the true coordinates of the circle's center are stored as a label.

python3 generate_synthetic_spheres.py --num_images 1000

python3 train.py --n_epochs 20 --batch_size 8 --save_as_onnx True

python3 visualize_results.py --model best_model.pth