search

donnyyou / torchcv

TorchCV: A PyTorch-Based Framework for Deep Learning in Computer Vision
https://pytorchcv.com
Apache License 2.0
2.25k stars 378 forks source link

Question about the implementation of the AlignModule in SFNet #157

Closed veizgyauzgyauz closed 4 years ago

veizgyauzgyauz commented 4 years ago

Hi! Thanks for sharing such a great project! As described in the paper, the low feature and the warped high feature will be sumed up in the Flow Alignment Module. But in the class AlignModule, the warped high feature is returned without summation. Could you tell me the reason? Besides, have you done any ablation study about the fusion strategy of the low feature and the warped high feature such as summation, multiplication or none? Thank you so much!

lxtGH commented 4 years ago

FAM is based on FPN design where summation is the default design. The main point is we improve the simple upsampling process with learned flow field.