Questions about the Good Semantics to localize experiment(Fig 10.)

[99angpaka]

Hello, Sarlin. your research is very intersting! Thank you for your good research.

However, I asked a question about the experimental design of "Good Semantics to localize".

As described in the "which map elements are most important? section in the Appendix, how did you experiment by removing objects?

Exactly, I want to know how you removed each element from the map and I also want to ask if it is included in the code.

Thank you.

[sarlinpe]

This figure was generated by running the MGL evaluation and successively masking out each of the different classes with the mask_index argument:

OrienterNet/maploc/evaluation/run.py

Line 59 in 38e60a1

mask_index: Optional[Tuple[int]] = None,

OrienterNet/maploc/evaluation/run.py

Lines 76 to 78 in 38e60a1

	if mask_index is not None:
	mask = batch["map"][0, mask_index[0]] == (mask_index[1] + 1)
	batch["map"][0, mask_index[0]][mask] = 0

It is a tuple (index_type, index_class) where index_type$\in\{0,1,2\}$ (areas, ways, or nodes), and index_class is the index of the class in the list of the corresponding type maploc.osm.parser.Groups.[areas|ways|nodes]:

OrienterNet/maploc/osm/parser.py

Lines 207 to 210 in 38e60a1

	class Groups:
	areas = list(Patterns.areas)
	ways = list(Patterns.ways)
	nodes = list(Patterns.nodes)

Results for each class can thus be obtained with:

from maploc.evaluation.mapillary import run

results = {}
for index_type, classes in enumerate((Groups.areas, Groups.ways, Groups.types)):
    for index_class, class_name in enumerate(classes):
        results[class_name] = run(..., mask_index=(index_type, index_class))