Trying out a thought.
I take image of a banana. I have a 3d mesh model of a banana.
Since there are semantic alignment methods, it is possible to match points. Once you have
3d-2d correspondence, in principle it should be possible to align the 3d mesh model to the imaged banana.
For now I hard-code the 3d-2d correspondence. See main.cpp
for the values.
I use Theia-sfm's DlsPnp method.
It works!
mkdir build
cmake ..
make
./main