In my experience, retinaface cannot precisely locate the 5 landmarks on rotated faces, especially for 90 degrees and more rotation. How to solve that problem without rotating the input image? Or can we find that the face is rotated thats why the landmarks are not good?
Not sure if you still need help but for anyone else ending up here : retraining/finetuning the model with rotation augmentation easily enables you to predict rotated faces
In my experience, retinaface cannot precisely locate the 5 landmarks on rotated faces, especially for 90 degrees and more rotation. How to solve that problem without rotating the input image? Or can we find that the face is rotated thats why the landmarks are not good?