A high-speed multi-GPU implementation of bottom-up attention using CUDA