Different machine learning frameworks in a C++ application

We want to use machine learning models and Python code written for TensorFlow, PyTorch and MXNet under Windows and Linux and demonstrate their use inside a C++ application. We can use open-source Python code and models for video object detection and activity recognition for this demonstration.

We need a common interface for video object detection/classification or activity recognition functions using which we can specify the ML framework, model or code to apply on a particular data. The underlying model or code should be invoked using the appropriate framework and the detection, classification or recognition results passed back to the calling function. We need a main command line application that demonstrates this common interface with the operating parameters (e.g., framework, model, code, data file) being specified in an XML file. The demonstration should work under both Windows and Linux. We should be able to use Python-based frameworks from the C++ application.

We want you to compile TensorFlow 2, LibTorch, MXNet, etc., create all the libraries from scratch, document the process and demonstrate the use of these frameworks in a sample application for video detection/classification and activity recognition (with and without GPU support). Deliveries will include the Visual Studio solution/Make files, all source codes and a short documentation.

Payment will be made in two milestones.