Real2Sim Image Adaptation

2019 - Course Project

Real2Sim

Project Description

Image domain adaptation through the conversion of images with randomized textures (or real images) to a canonical image representation. This project was a replication of a RCAN paper with different loss modeling.

Instead of using GAN loss, this implementation used Perceptual/Feature Loss for better training stability and results.

Technical Approach

  • • Domain adaptation from randomized textures to canonical representations
  • • Implementation of Residual Channel Attention Networks (RCAN)
  • • Perceptual loss instead of adversarial training
  • • Focus on sim-to-sim transfer for robotics applications