What is a proper format of labels?

[tugot17]

I have a custom dataset. 4 Classes: ['Port', 'Necklace', 'Endo', 'ICD']

I have a dataframe where each row represents a separate detection. If some image has more then single detection, then it's image_id is in two columns

image_id	label	width	height	bbox	path
00005885_000.png	Port	1024	1024	[262.0, 355.0, 365.0, 435.0]	/home/piotr/Desktop/Nih-Chest-X-Ray-Artifacts ....
0005776_009.png	Necklace	1024	1024	[386.0, 1.0, 696.0, 216.0]	/home/piotr/Desktop/Nih-Chest-X-Ray-Artifacts ....

and I wrote a custom Parser

class ArtifactParser(DefaultImageInfoParser, FasterRCNNParser):
    def __init__(self, df):
        self.df = df
        self.imageid_map = IDMap()
        self.labelid_map = IDMap()

    def __iter__(self):
        yield from self.df.itertuples()

    def __len__(self):
        return len(self.df)

    def imageid(self, o) -> int:
        return self.imageid_map[o.image_id]

    def filepath(self, o) -> Union[str, Path]:
        return o.img_path

    def height(self, o) -> int:
        return o.height

    def width(self, o) -> int:
        return o.width

    def labels(self, o) -> List[int]:
        return [self.labelid_map[o.label]]

    def bboxes(self, o) -> List[BBox]:
        return [BBox.from_xyxy(*json.loads(o.bbox))]

Everything seems to be working but I'm not sure whether it is a proper approach for creating a Parser or should I change something?

[lgvaz]

For labels you should use:

self.class_map = ClassMap(['Port', 'Necklace', 'Endo', 'ICD'])

And then at the imageid method:

return [self.class_map.get_name(o.label)]

We still have to better document this, but ClassMap will take care of adding the background class at index 0, which is required by the models we currently support.

On a quick look everything else seems fine on the parser =)

---

Note: Very clever to load the bboxes with json.loads, much easier than what I'm doing in the wheat parser ^^

[tugot17]

I changed as you recommended to self.class_map = ClassMap(['Port', 'Necklace', 'Endo', 'ICD']) and it works perfectly, however in show_record I still need to give class_map= as an argument.
Maybe it would be a good idea to use the field class_map from ArtifactParser class (eg. make it a Parser argument) in show_record method? By doing so there would no more be a necessary to use this variable each time when we want to display the image?

[lgvaz]

There are a couple of issues with that,

First, you only pass the record to show_record, so we would need to add a reference to class_map in each record.

Second, class_map is not a required attribute for the parser, take note that you had to define it yourself, this is by design, some datasets already give you the label_id instead of the label name, and there is no reason to use a class_map in those situations.

In summary, the way it's implemented gives more freedom to the user, you're not forced to use ClassMap if you don't want to, it's just a convenience class.

[tugot17]

Okk, makes sense. So, for now I close the issue. Thanks for your help :)