C++ Acceleration and Modern Architectures for {torchvision} in R

The {torchvision} library in mlverse is already near-feature-parity with the PyTorch library, but certain performance optimisation and state-of-the-art computer vision model support gaps need to be bridged. This project aims to bridge the performance optimisation and state-of-the-art computer vision model support gaps in the current mlverse library. The performance optimisation objective is to improve the performance of the library by moving computationally intensive functions, e.g., region proposal networks and non-maximum suppression, from R to C++. This will allow for seamless integration with libtorch, providing production-grade speed and performance for detection and segmentation tasks. The state-of-the-art computer vision model objective is to improve the library by adding YOLO, RT-DETR, and SAM models, among others, into the torchvision library. This will be achieved by providing end-to-end documentation of the models, e.g., using the R documentation system, roxygen. This project will result in improved performance by providing an R interface for performance-critical functions using Rcpp, as well as improved state-of-the-art computer vision models by adding YOLO, RT-DETR, and SAM models, among others, into the torchvision library. This will improve the performance, usability, and adoption of R as a deep learning library for computer vision tasks, making mlverse a competitive alternative