Semantic segmentation evaluation - ignoring background

[anmolkalia]

In the implementation of miou for the semantic segmentation evaluator:
I want to understand why pos_gt is not calculated as
pos_gt = np.sum(self._conf_matrix[:, :-1], axis=0).astype(np.float)
instead of https://github.com/facebookresearch/detectron2/blob/main/detectron2/evaluation/sem_seg_evaluation.py#L133
In the current implementation, the pixels for the semantic class that are predicted as background are ignored.

Similarly for pos_pred, by doing https://github.com/facebookresearch/detectron2/blob/main/detectron2/evaluation/sem_seg_evaluation.py#L135, we're ignoring the pixels for which some semantic class was predicted but the gt class is background. Should this not be the following instead:
pos_pred = np.sum(self._conf_matrix[:-1, :], axis=1).astype(np.float)

Is my understanding incorrect?