`label_map` used twice during evaluation

[panfeng-hover]

I write a customized dataset, where the label ids are (0, 1, 2, 3, 4, 5, 6), and I would like to merge some classes, so I set

1. self.custom_classes = True
2. self.label_map ={4: 3, 6: 3, 5: 4}.
   Namely convert label ids of 4 & 6 in the ground truth to 3, and convert label id 5 to 4, as mmsegmentation is assuming the label ids is continuous,

However, it looks like the label_map is wrongly used during evaluation, in details,

the self.get_gt_seg_map_by_idx(index) used here already converted the ground truth label ids from (0, 1, 2, 3, 4, 5, 6) to (0, 1, 2, 3, 4), but during later intersect_and_union, the label_map is used again, so the ground truth label ids will become (0, 1, 2, 3) from (0, 1, 2, 3, 4).

[panfeng-hover]

@MengzhangLI would u like to take a look? Thanks!

[panfeng-hover]

It is similar to the one used in evaluate function, where gt_seg_maps = self.get_gt_seg_maps() is already converted by the label_map during pre_pipeline