To assess the performance of the program we use:
-
mean Intersection over Union (mIoU): $$ mIoU = \frac{1}{number; of; images; evaluated} \cdot \frac{Area; of; intersection; of; two; boxes}{Area; of; union; of; two; boxes} $$
-
Accuracy: number of instances correctly recognized for each object category. Correct prediction if
$IoU > 0.5$