Two modules in the end-to-end AdMiRe architectures are super resolution and background extraction applications. Both modules follow a similar internal, are based on deep learning techniques, and are implemented on a standalone workstation enhanced by a GPU that receives compressed video captured and streamed by a remote participant and feeds the next stage of the workflow with a compressed video of the same format. The modules can be used in three different ways:
1- If the remote user possesses a green screening set up, the background extraction module can be bypassed.
2- If the resolution of the camera used by the end user is high (e.g. a 4K webcam), the super resolution module can be bypassed.
3- In all other cases, the operator can choose to turn on, one or both modules.

By the end of the first year, of the AdMire project, in August 2021, an implementation of both modules in docker was produced and shared with AdMire projet partners for the purpose of integration and testing. EPFL, the partner in charge of the design and implementation of these two modules, plans to release them in form of open source after their final optimization and validation, in order to allow their use in a larger range of applications, beyond those under the scope of the AdMiRe project.

The following activities are under way in order to finalize the deliverables containing the two modules:
–       Assessment of the impact of video format and compression parameters in the quality of super resolution and background extraction
–       Performance of the background extraction in scenarios where subjects manipulate and showcase various types of objects in different sizes and shapes, including those made of transparent materials.
–       Performance of the super resolution and background extraction modules in various environments with different types of set up configurations and different illumination conditions
–       Comparison of background extraction module with green screening in different scenarios
–       Recommendation in form of best practices for remote participants as well as operators, to allow them optimization of the quality of the final result after super resolution and/or background extraction modules.

Touradj Ebrahimi / EPFL