Overview

This project explores the power of image-to-image and video-to-video translation using advanced machine learning techniques. We implemented Pix2Pix and CycleGAN models to perform satellite-to-map translation, facade segmentation, and artistic style transfer on videos.

Final Results

Satellite-to-Map Generation

Artistic Image Style Transfers

Videos

Watch the output of our models in action:

Methodology

Our process involved breaking large images into smaller patches for processing, applying advanced blending techniques to stitch results, and optimizing video frames for high-quality output. Key steps included:

• Using Pix2Pix and CycleGAN for image-to-image and video-to-video translation.
• Preprocessing frames with techniques like histogram matching, Gaussian-Laplacian pyramid blending, and denoising to improve video quality.
• Stitching processed image patches to form cohesive maps or videos.

Results

Our results demonstrate the effectiveness of CycleGAN over Pix2Pix in generating higher-quality outputs with less noise. For artistic style transfer, CycleGAN maintained object structure better, requiring fewer additional processing techniques. Check out the captivating results in our final processed videos and maps!

Final Report

Read our comprehensive report on this project: