HAttMatting vs fba, gca, etc

[javiermachinelearning]

good day and congratulations for your work :)

• your paper says that results may be slightly worse than 2 other models : "Ou HAttMatting is slightly inferior to Context-Aware and IndexNet.",

• however those 2 require intermediate trimaps: "Although they both generate high-quality alpha
  mattes, trimaps are strongly required during their training and inference phase, which restricts their effectiveness in practical applications. Our HAttMatting only need single RGB images as input, which is very convenient for novice users.",

• I would say that not needing trimaps is not only better for novice users, its just better all around, as perfect trimaps are difficult to automatically create and trimaps directly impact the quality of the matting, so that´s great, but now, a possible concern for me is that if you look at the ranking table at
  http://alphamatting.com/eval_25.php

"Context Aware matting" is in position 18 of the table, way below FBA and GCA matting (fba and gca are the two deep learning trimap-using approaches I have been working with already).

• This seems to hint about end to end approaches (like late fusion and HattMatting) having maybe worse results than trimap-using approaches (like FBA and GCA) ? is that correct?. The question I guess is how much worse, if its just a bit worse then its much better to use an end to end approach like HattMatting.

• any tips? I have been working with FBA and GCA which are good in that table http://alphamatting.com/eval_25.php, but I would much prefer to use an end to end model like HAttMatting,
  however the question is how much worse, if its worse, are the results of HAttMatting compared to FBA or GCA I wonder

• for my objective which is to extract people from pictures, people in groups, full body, portraits etc, so extract them well with hair, etc from pictures, (I am planning to create a new dataset probably to train the model, a complete and diverse dataset)

thank you for any tips :)