csxmli2016 / DFDNet

Blind Face Restoration via Deep Multi-scale Component Dictionaries (ECCV 2020)
915 stars 215 forks source link

Cropping posibilties #8

Closed Lubsey1 closed 4 years ago

Lubsey1 commented 4 years ago

Hi, is it a possibility to process a face in an entire frame (1920*1080) rather than cropping and scaling it to just the face?

Thanks.

csxmli2016 commented 4 years ago

Hi, is it a possibility to process a face in an entire frame (1920*1080) rather than cropping and scaling it to just the face?

Thanks.

Hi, this may be a valuable topic. The simple way is to copy and paste the face results to the original entire frame with upsample scale s, which can only enhance the face region. This is what I am doing (refer to [here])(https://github.com/csxmli2016/DFDNet/issues/6#issuecomment-672392265).

Another complex way to process a face in an entire frame should flexibly explore the face component in different sizes and locations in CNN. But this may be limited in our current framework due to the requirement of dictionaries in the same feature space.

csxmli2016 commented 4 years ago

Hi, is it a possibility to process a face in an entire frame (1920*1080) rather than cropping and scaling it to just the face?

Thanks.

Hi, I have updated a new branch. It is more flexible to handle an entire frame. The background will be enhanced in my future work.