Project information

  • Category: Style Transfer | Computer Vision with Patch Segmentation and Extraction
  • Purpose: Durham University
  • Assignment Grade: 70%
  • Project date: May, 2022

GTA Style Transfer Coursework

This coursework performing a style transfer between old movie footage and GTA gameplay. A Mask-RCNN was used for human extraction and OpenPose was used for pose estimation (e.g. full-body, half-body, head-only, etc.). Each category of patch was then augmented to prepare for training and then inputted into separate CycleGANs to perform the style transfer. Resultant images were then overlaid onto the original frames to reconstruct the video with the new style-transferred humans. Code can be provided upon request.

Final Output:

Open on desktop to view PDF