I am trying to implement 3D versions of NMS and RoiAlign for a 3D image detection network by taking some C++ and CUDA code from different sources and modifying them for 3D image volumes with BCHWD sizes. One thing I am curious about (and it isn’t clear in the documentation) is if the current NMS and RoiAlign code are meant for normalized or un-normalized coordinates for their box inputs. This won’t really be an issue for the NMS (since the intersection and union calculations would both be normalized) but may have an impact on the RoiAlign fuction.