search

zorzi-s / PolyWorldPretrainedNetwork

PolyWorld: Polygonal Building Extraction with Graph Neural Networks in Satellite Images
Other
175 stars 28 forks source link

Problem with choosing backbone #4

Closed 654493176 closed 2 years ago

654493176 commented 2 years ago

Hi，thank you for your excellent work and code! I notice you used R2UNet as backbone in your network, compared with the Frame Field Learning which used UResNet101. And in this case, your algorithm is more effective. I'd like to know have you tried to use other backbone(like UResNet101 same as FFL) in your network? will them show similar results? I'll very appreciate for your reply.

zorzi-s commented 2 years ago

Hi @654493176,

thank you for the interest on my work! I didn't deepen into analyzing the performance of different backbones for this task, so I can't really tell for sure which one is better. I can say that I made some early experiments with the UResNet101 (proposed by FFL) and I noticed almost equivalent performance to the residual U-Net.

654493176 commented 2 years ago

Thank you for your reply. I think if your algorithm don't rely on a particular backbone, you should use same backbone in all network when you compared them(It's easily done, FFL has complete train code). Otherwise,we can't be sure that better results come from different backbone or your excellent algorithm. Sorry about my poor English，I wish you can understand what I mean.I do not doubt your result,I just think they need a fair Comparison ,Actually in my experiments,R2UNet underperforms UResNet101 on CrowdAI dataset. Good luck with your research career!

zorzi-s commented 2 years ago

@654493176 Thank you for the comment and the suggestion. I will consider to compare different backbones if an extended version of the paper will be released.

654493176 commented 2 years ago

Thank you for your reply and I have no other questions.Looking forward to your further results.