Confusion Matrix & Mean IoU

The problem, from first principles

The standard score for semantic segmentation is mean intersection over union. For each class, take the pixels predicted as that class and the pixels that truly belong to it; IoU is the size of their overlap divided by the size of their union. Averaging over classes gives mIoU. Everything is read off a confusion matrix whose entry C[i, j] counts pixels with true class i predicted as class j: the diagonal holds correct pixels, a row sum is the true size of a class and a column sum is its predicted size. A class that never occurs and is never predicted has an undefined IoU (0/0) and is excluded from the mean.

From theory to code

Implement confusion_matrix and mean_iou.

Constraints

• confusion_matrix(pred, target, n_classes) takes integer arrays of equal shape with values in 0..n_classes-1 and returns an (n_classes, n_classes) integer matrix with C[true, pred] counts.
• mean_iou(conf) computes per-class IoU_c = C[c, c] / (row_sum_c + col_sum_c - C[c, c]) and returns (miou, ious), where classes with a zero denominator get nan in ious and are ignored by the mean (np.nanmean).

Hints