Physics-Based Synthesis

← Back to Synthetic Data Generation

Physics-based synthesis places parameterized particle models onto real or procedurally generated backgrounds. Because the object positions and shapes are known by construction, ground-truth segmentation masks are automatically available.

Typical workflow:

1. Acquire a set of representative background images (empty substrate, binder phase, etc.).

2. Define a particle model: shape (circle, ellipse, polygon), size distribution, contrast, edge-blur radius.

3. Randomly sample particle parameters from the defined distributions.

4. Composite particles onto backgrounds using alpha blending.

5. Record the binary mask of each particle.

This approach is related to domain randomization [13], which deliberately over-randomizes simulation parameters to force the trained model to generalize across a broad range of conditions.

See also

• Classical Augmentation — a transformation-based alternative that doesn’t need a particle model

• Deep Convolutional GAN (DCGAN) — a learned, data-driven alternative to hand-specified particle models

• Synthetic Data Generation — every other synthetic data strategy