Segmenting object parts such as cup handles and animal bodies is important in
many real-world applications but requires more annotation effort. The largest
dataset nowadays contains merely two hundred object categories, implying the
difficulty to scale up part segmentation to an uncons