Thank you for your awesome work on this project and for making the code and training data available!
I had a question regarding the real train dataset - as a sanity check, I've been visually ensuring that the virtual 3D bounding boxes align well with the drill object. From your paper, I gathered that you used ICP for estimation of the ground truth drill pose. I would expect that this could introduce some error in the ground truth pose relative to the "actual" drill pose and wanted to ensure that the error that I'm experiencing is similar to what's expected. I've included a few sample images below with their respective bounding boxes drawn, do you think that the accuracy of ground truth placement looks reasonable in these examples?
Good alignment
Poor alignment - looks like mostly rotational error introduced, more visible at instrument tip