Question about the logic of processing parent prediction in visualization_parts()

[SevenLJY]

Hi,

Thanks for the amazing work.

I'm trying to understand the output processing code in the demo and get confused about how you get the final parent node for each part. Specifically, I didn't get the code at lines 443-447 in util.py as shown below.

# update parent labels
for p_i, each_parent in enumerate(parent_pred):
    if each_parent>0:
        if each_parent-1 in list(new_order):
            parent_pred[p_i] = list(new_order).index(each_parent-1)+1

Can you explain a bit of the intuition behind how you update the parent labels? Many thanks.

[qiuyuchen14]

Hi Jiayi, I'm really sorry for the late reply and thanks so much for your interest! Here each_parent represents the parent's index in the list of links. if each_parent=0, that means it's parent is the root (for example the cabinet frame). I define all the doors, drawers etc in the first level that are directly connected to the root (each_parent=0). The code here is trying to update the parent index of the second-level parts such as handles and knobs (each_parent>0), so that their predicted parent index are matching the first-level parts in the new_order. The new_order here contained the reorder level index so that first-levels go first. This is because the order of the bbox (either from the detection or you manually labeled) might not come in the order of the tree structure (first door, then handle). Hope this anwsers your questions!