It takes as long as it takes. Using multiple dispatch queues is unlikely to help, since you're limited by how fast the hardware can go. If you want a faster body pose detection method, you will have to train your own model.

Does your model use a NonMaximumSuppression layer (or as part of a pipeline)?