About topo heads

[eeeric1]

Hi, thanks for your work. I am a little confused about the topo head code, and I would like to ask you for advice. In line 133 of topo_lt_head, targets need to be operated according to the shape of xs. I'm not quite sure why this is done. I understand that we can do the same thing as line 157 of topo_ll_head. Looking forward to your reply!

[wudongming97]

Hi, It means computing the LT loss for only positive samples. It can increase the loss weights of positive samples (they are not too much in all samples).

[eeeric1]

Thanks for your reply! I understand the purpose of this, but I'm not sure why topo_lt_head and topo_ll_head do it differently. in topo_lt_head.py

if xs.size(0) != xs.size(1):
     targets = 1 - targets[0][xs_new].view(-1).long()
else:
     targets = 1 - targets[0][xs_new][:, ys_new].view(-1).long()

in topo_ll_head.py

targets = 1 - targets[0][xs_new][:, ys_new].view(-1).long()

[wudongming97]

From my experience, the ll has more positive samples. Because each centerline has more than one connected centerline, most centerlines may not connect any traffic element.

[eeeric1]

Hi, I'm still a bit confused about the code here The main problem lies in this code

if xs.size(0) != xs.size(1):

In my opinion, you can directly take targets[0][xs_new][:, ys_new], which is consistent with topo_ll_head, or in order to add some samples, take targets[0][xs_new]. But I don’t know why xs.shape is used to determine how to get the target.

[wudongming97]

Oh. Sorry for that. In our initial code, the ll and lt heads are not separated, so use this judgment. It is not important.

[Wolfybox]

Thanks for your reply! I understand the purpose of this, but I'm not sure why topo_lt_head and topo_ll_head do it differently. in topo_lt_head.py

if xs.size(0) != xs.size(1):
     targets = 1 - targets[0][xs_new].view(-1).long()
else:
     targets = 1 - targets[0][xs_new][:, ys_new].view(-1).long()

in topo_ll_head.py

targets = 1 - targets[0][xs_new][:, ys_new].view(-1).long()

@wudongming97 I confused why there is a inverse operation here to make pos to 0 and neg to 1:

targets = 1 - targets[0][xs_new][:, ys_new].view(-1).long()