Rethinking archive PostProduction with LEarning, vAriational, and Patch-based methods – PostProdLEAP

The goal of the PsotProdLEAP project is to develop new tools for video archive post-production by leveraging both recent deep learning approaches and patch-based and variational approaches. Frequent artefacts observed with deep-learning methods include loss of details, spatial and temporal discontinuities, and colour bleeding along edges. These artefacts make state-of the-art deep learning approaches inappropriate for professional post-production. In the framework of the PostProdLEAP project, these limitations will be solved by designing deep learning models including spatial and temporal regularization and constraints on texture features. These tools will be developed in collaboration with the artists and historians from Composite Films, world leader in restoration and colorization of films and archive footage. Through this collaboration, we will enable artists interaction and give them control on the final results, while making their tasks less time consuming and tedious than with current professional softwares. Our models will be trained, tested and validated on datasets of video archives with different resolutions and quality, created specially for the project from movies restored by Composite Films.