This repo contains variation methods of colourisation.
Overview of methods and their relation to the project
| Method | Work Package | Status Comments | UI (Python) |
|---|---|---|---|
| Deoldify | 1.2 (No user input) | Working with quality improvements, artistic style net not fully tested | Need to select which model (artistic/stable). Scaling parameter included |
| Fully Automatic Video Colorization with Self-Regularization and Diversity | 1.2 (Colour Net) , 3.5 (TC Net) | Author emailed the TC Net weights but not yet implemented | Not yet |
| Interactive deep colourisation | 1.2 (No user input) , 1.4 (Global colour Net) , 2.2 (Local pixel Net) | Global Net is in caffe with custom compilation that is fiddly to set up, have emailed authors and no pytorch release planned | Local Net... need make UI for uploading colour image and selecting which areas to change colour |
| Colorisation using Optimisation | 2.2 | Code runs slowly --- need heavy optimised code to gain performance | Need same as above |
| Deep exemplar based video colourisation | 1.3 (single image reference) , 1.2 (given multiple stock references) , 3.1 , 3.2 (given multiple stock references) , 3.3 (single image reference) , 3.5 (TC loss function for future network training) | No code given so implementing method from scratch, some mistakes in the paper that the autor is not replying about | - |
| Image segmentation | 2.1 | Basic look into methods - seems very doable | - |
| Tracking emerges via colourisation | 3.1 | Network is low quality - could tweak network to make better quality video output (U-Net) | - |
Use the shell scripts inside set_up to install everything. Currently tested on Ubuntu 18.04. Make sure computer has an advanced vector chipset architecture for use with tensorflow (any intel CPU post around 2010) or you'll have to compile tensorflow from source otherwise. Also a decent nvidia GPU is necessary.
Once everything has been set up, enter python virtual environment and install colourisation as an editable package
pip3 install -e .