Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

front_x or cross_x? #4

Open
sunnyHelen opened this issue Sep 28, 2021 · 1 comment
Open

front_x or cross_x? #4

sunnyHelen opened this issue Sep 28, 2021 · 1 comment

Comments

@sunnyHelen
Copy link

front_res = torch.cat((front_x, T), dim=1)

output = front_x + front_res

Hi, I was confused about the calculation of cross-view transformer. According to the Figure 4 in your paper, it seems that the "front_x" should be "cross_x" in Line 52 and 55. Am I misunderstanding anything? Please help me figure it out.
Looking forward to your reply and thanks!

@ChristopheZhao
Copy link

ChristopheZhao commented May 8, 2022

I have test the difference between front_x and cross_x that concanate with T:
sample 1(cross_x is better):
origin front:
000707
the res used front_x:
000707
the res used cross_x:
000707

sample 2(cross_x is better):
origin front:
001477
the res used front_x:
001477
the res used corss_x:
001477

sample 3(front_x is better):
004078
the res used front_x:
004078
the res used cross_x:
004078

And more 10 sample i tested has same degree result that i cant distinguishing their difference,so i confused if the pretrain model used front_x concanated T to train data.But maybe the right ways is to concanated with cross_x,So although use cross_x concanated with T to test sample,also can achive the same result as font_x.

# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants