Our vision pipeline uses Argus and the NvBuffer API (nvbuf_utils.h) to handle format conversions, resizing, etc. My next goal is to add a segNet from jetson-inference to our processing, which takes floating-point RGBA.
I plan on using the jetson-utils cudaNV12ToRGBA32 kernel with cuGraphicsEGLRegisterImage to convert from NvBufferColorFormat_NV12 to floating-point RGBA, but jetson-utils provides no kernels for going the other direction.
Is there any good way to handle floating-point RGBA buffers with nvbuf_utils or should I use a kernel to convert back to something like NvBufferColorFormat_ARGB32 (ARGB-8-8-8-8) and go from there?
Thanks!