Parametrize number of semantic labels

[mit-spark]

Parametrize number of semantic labels

[ToniRV]

Hi! That sounds like a great idea! I'll get back to you later with suggestions 👍
The idea of fixed Eigen types was just to make it faster, but dynamic matrices should also work.

[codieboomboom]

@RozDavid I realized the same thing today when working on a different segmentation model trained on ADE20K. Seems like the value is defined in this header file for kimera_semantics module.

Something i observed for my side is when changing that constant to my model's number of label, running the kimera semantics will fail this assertion code in semantic_integrator_base.cpp. Not sure if you know why it happens, I just disabled it for now.

CHECK_NEAR(
      semantic_log_likelihood_.sum(),
      kTotalNumberOfLabels * log_match_probability_ +
          std::pow(kTotalNumberOfLabels, 2) * log_non_match_probability_ -
          kTotalNumberOfLabels * log_non_match_probability_,
      1000 * vxb::kFloatEpsilon);

[RozDavid]

Hey @AnhTuDo1998,

Yeah, I started to work on implementing the dynamic number of labels yesterday, and met with the same problem when we set the total number over 25ish. It is only a rounding error I believe, so we just need to adapt the treshold with the number of labels there, or just manually set it a megnitude larger. For sanity check I think it will still stand.

I believe the bigger challenge here is the initialization of SemanticVoxel priors when assigning new Voxblox blocks with a constant matrix, that's size and value depends on the parameter. With non-fixed sized Eigen matrices we cannot use the original Structs that was defined here.

[ToniRV]

I added my comments to #56, that is definitely the way to go!