"Person" class in semantic segmentation input

[junweiliang]

"Person" class in semantic segmentation input

[JunweiLiang]

Originally in the next-prediction repo, we use ADE20K model and only these most common 10 classes are used in preprocessing, which does not include "person" class since person features are covered in another module. See here. Since in the ActEV dataset these are the most common classes except for person class and to keep the same preprocessing, I put the person class into the background (id 0). This is a bit hacky. You can easily change this by adding an entry in the scene36_64_id2name_top10.json file ("13" is person class for ADE20K) for training and testing but may not see a performance difference.

[TheDarkKnight1998]

Thank you for your explanation！