What do we consider an ID switch?

[sfmig]

• Before: we computed symmetric difference (aka, total number of elements that are not in common)

• @nikk-nikaznan made some corrections to this in PR MOTA revisited  #181

  • we were counting re-IDing twice?

• right now we cannot detect when two crabs that are close together and swap IDs from frame f to f+1 (with no IDs created)

  • do we want to detect this?
  • likely we need to check bboxes overlap between detection and GT

• should we split ID switches into categories?

  • be careful about counting twice
  • is it worth it?

[sfmig]

After a long and very helpful discussion we are clear on our definition, but it's probably a good idea to put it in writing - that is covered in #214