Check if surface normals are mirror-reversed in the ultrasound plane #8
Description
Some objects, such as the mustard bottle, soup can, and Harissa-oil jar, have noisy surface normals, even after the various refinements to the data-collection pipeline. While these could be due to persistent noise in the probe pose estimate, etc., one possibility is that the surface normals are being accidentally mirror-reversed in the imaging plane during the processing pipeline. That is to say, if the surface normal is detected as pointing slightly left, it is projected into the coordinate space of the learned object in the opposite direction. If this was happening, it would not be apparent for all observations, because when the surface normal is directly parallel to the axis of the probe (i.e. pointing directly at the tip of the probe - quite a common situation due to the scanning technique employed), then no offset in their axes is present. Instead, the constraints of the scanning movement (such as certain regions of the fluid bag / object only being accessible at a certain angle) would have resulted in clusters of these surface normals appearing. Images below show examples of odd surface normals that could fit this issue.
The task is to investigate whether the code is consistent with an incorrect orientation of the surface normal direction, and whether reversing this results in better learned models with surface normals that are more consistent with our prior expectations. Implementing a fix (if necessary) may also result in improved generalization, due to better alignment in observations between different datasets (sim to real or real to real).
Below are some example objects, highlighting the kind of surface normals whose direction appears consistently biased in an inappropriate direction.
Mustard bottle:
Soup can:
Harissa oil jar:
