I am still not sure why you don't solve it in post processing, but one alternative way would be to make two different predictions. The first gives the location of object 1, the second provides the distance/angle to object 2.