Realistic Shadow Generation with GAN

This project demonstrates realistic shadow generation on natural images using a custom-trained GAN model. Given a deshadowed input image, the model learns to cast shadows in appropriate regions, conditioned on the shadow mask and instance segmentation.

---

Project Structure

• 1_shadow_generation_code.ipynb: Main Colab notebook for model inference
• assets/: Folder containing input, output, and mask visualizations
• models/: Generator and discriminator architectures
• utils/: Image preprocessing and visualization functions

---

How It Works

1. Input: A deshadowed RGB image
2. Preprocessing: Generates:
   • Shadow mask (binary)
   • Instance mask (semantic segmentation)
3. GAN-based Generation: Produces a realistic shadow-cast image
4. Post-processing: Visualization and export

---

Technologies Used

• Python, PyTorch, NumPy, OpenCV
• Google Colab for experimentation
• GANs: U-Net Generator, PatchGAN Discriminator

---

Notes

• This is a research-oriented prototype for cast shadow synthesis.
• Ideal for use in training data augmentation, simulation, or visual effects.

---

Acknowledgement

Sample image of zebra sourced from open dataset used for testing.