- James Dixon
- Robert Ace Gonzales
- Steve Ma
- Arye Oskotsky
- Course: CSC 671 Deep Learning
- Instructor: Dr. Robert Mateescu
- Institution: San Francisco State University
- Date: October 10, 2024
This project aims to apply Style Transfer to spherical (360-degree) images using Convolutional Neural Networks (CNNs). Spherical images are typically represented as equirectangular projections, which pose challenges for traditional style transfer methods due to:
- Distortion caused by flattening spherical images.
- Difficulty in handling transitional edges.
- Develop a method to perform style transfer while minimizing distortions and artifacts, particularly around the seams of the image.
- Leverage rectilinear projections to maintain seamless transitions across the entire spherical surface.
- Source: Microsoft COCO dataset
- Implement a CNN for style transfer on ordinary (non-spherical) images.
- Based on the work of Gatys, Ecker, and Bethge (2016):
Image Style Transfer Using Convolutional Neural Networks
- Extend the CNN from Phase 1 to work on spherical images.
- Approach inspired by Ruder, Dosovitskiy, and Brox (2018):
Artistic Style Transfer for Videos and Spherical Images
- Perform hyperparameter search to optimize the network for spherical style transfer.
- Focus on minimizing artifacts, particularly near the seams of rectilinear projections.
-
Gatys, L. A., Ecker, A. S., & Bethge, M. (2016). Image Style Transfer Using Convolutional Neural Networks.
DOI: 10.1109/CVPR.2016.265 -
Ruder, M., Dosovitskiy, A., & Brox, T. (2018). Artistic Style Transfer for Videos and Spherical Images.
DOI: 10.1007/s11263-018-1089-z
- Clone the repository:
git clone https://github.com/JD499/style-transfer.git cd style-transfer