Description
Hello. According to the comment:
# The case where n_true_groups == n_pred_groups == 1 needs to be
# special-cased (to return 1) as the above formula gives a divide-by-zero.
# This might not work when true_mask has values that do not sum to one:
I understand both_single_cluster is handling this corner case.
What I am confused is this function: https://github.com/deepmind/multi_object_datasets/blob/master/segmentation_metrics.py#L97. It seems to me it makes sense when it is applied to ground truth masks as _all_equal(true_group_ids) , but when it is applied to _all_equal(pred_group_ids), the prediction slot for object may not necessarily be "the first slot" as suggested by values[..., :1].
May I ask if it is corner case missing in this implementation?
Thank you!
More specifically, your code first finds the argmax of the tensor with shape [bs, n_pixels, n_classes] over the last axis. Assume n_classes == 3, then because my prediction for object is in slot 3, I believe that argmax will return two values, 3 for the pixels with objects and 1 for pixels exclude the object, as argmax returns 1 by default if all values are equal. It seems to me that this will lead to a wrong condition in your ARI value.