In this project, we aim to recognize the shape of a 3D object based on an image of its shadow. The idea was inspired by the video "The average area of a shadow" by 3Blue1Brown.
This project explores multiple synthetic 3D shapes and uses machine learning models, including logistic regression and convolutional neural networks (CNNs) among others, to classify objects based on their shadows.
- Liam Farhan (Formerly Mohamad Al Farhan)
- Emil Reiter
- Mahmoud Sharaf
We created a dataset to train the models. The shapes are all synthetic.
- 3D Shapes: We generated 3D meshes for several shapes, including:
- Cube
- Pyramid
- Tetrahedron
- Diamond
- Dodecahedron
- Icosahedron
- Torus
- Cylinder
- Cow
- Human
- Projection: We used perspective projection to create shadows. In this method, the light source is replaced by a viewer, and the objects are projected onto a screen.
- For each object, we generated hundreds of shadow images by randomly orienting the object and projecting it onto the screen.
- Preprocessing:
- Edge detection was applied by comparing pixel values to identify object boundaries and generate images containing only the edges.
- Data was compressed and cast into boolean values. To handle memory limitations, all images were resized to 200x200.
- Training: The model was trained with L2 regularization.
- Accuracy: Achieved a classification accuracy of approximately 60%, which improved slightly (to 62%) when using edge-detected data.
- Performance: The model performed better on full shapes than edge-detected data. Using dropout regularization, a learning rate of
0.00001, and the Adam optimizer, the best results were achieved with a batch size of 40.
- CNN architectures, such as VGG and LeNet, were used with varying configurations.
- Results:
- The highest accuracy of 97.58% was achieved using a Conv-Conv-Pool architecture.
- CNNs proved to be superior to DNNs for this task.
