Testing with real-life images returns bad results

[willswordh]

@24wenjie-li
Hi, I try to test real-life images based on your model but the results are quite distorted. I wonder if I have to preprocess the images and crop the face by face landmarks to output good results from your model like the ones in your paper. Please help me with it. Thanks a lot!

[24wenjie-li]

Hi, thanks for your concern, this phenomenon is normal. Firstly, faces do have to be cropped plus aligned, but this is not the main cause of the distortion. The training dataset is the main cause of distortion. This article is mainly intended to explore efficient backbobe for FSR. To simplify the training, the low-resolution training datasets used by all methods are high-resolution datasets that have been downsampled, not complex real-world datasets or blind datasets. For good results on real images, it is recommended to test our methods after retraining on real/blind datasets. willswordh ***@***.***> 于2024年10月12日周六 07:47写道：

…

@24wenjie-li <https://github.com/24wenjie-li> Hi, I try to test real-life images based on your model but the results are quite distorted. I wonder if I have to preprocess the images and crop the face by face landmarks to output good results from your model like the ones in your paper. Please help me with it. Thanks a lot! — Reply to this email directly, view it on GitHub <#4>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ATEXBSUR77LGTUNM4GDPIDDZ3BPSFAVCNFSM6AAAAABPZZ6V2WVHI2DSMVQWIX3LMV43ASLTON2WKOZSGU4DEMZUGM4TOOI> . You are receiving this because you were mentioned.Message ID: ***@***.***>