MartinSmeyer
commented
2 years ago Hi Jianyu,
The first goal concerning symmetries is to avoid the ambiguities during training and avoid learning wrong averaged solutions. As a consequence codes of symmetric views are close/identical after training. For the evaluation we still predict one single pose because that is how approaches are typically evaluated.
But you can easily get multiple potential poses of a potentially symmetric view by returning, e.g. the 10 nearest codes by cosine similarity and then clustering them in pose space.
For axis symmetries you will get a pose distribution around one axis and you can find this axis by transforming the relative pose estimates into the axis-angle representation and then cluster/average the axes.
Best,
Martin
JZhao12
Hi Martin,
Sorry to disturb you, may I ask how did you solve the problem with symmetries? I read your paper and I found that the symmetry problem could be ignored during training the AAE, but it still exists in the codebook when comparing the cosine similarity.
If you have mentioned the solution in your paper, could you please point out what page it is on? Maybe I didn't read it carefully enough, sorry to bother you.
Thank you very much! Jianyu