Help calculating mIOU for segmentation metric

[cnavarrete]

I know the compute_ap function gives you the overlaps of the masks for a image, but I would like to know how could I use this to calculate the mIOU over the whole test dataset. Also I would be grateful if someone could explain what the overlaps array actually means

[[1.0454647 e-04 9.3533748 e-01 0.0000000e+00]
 [9.4103068 e-01 3.4195228 e-04 0.0000000e+00]
 [0.0000000 e+00 0.0000000 e+00 9.3462163e-01]]

I have trouble understanding this, why is returning a 3x3 matrix, I though it would be something like the mean of the overlaps for the image or something like that.