Retraining 512x512 with 68 keypoints，then retraining 512x512 with dlib 68 keypoints detector

[Qia98]

I want to retrain FOMM with 68 keypoints (512x512) and release it publicly, if successful. In addition, I prefer to retrain fomm based on dlib 68 keypoints detector. This allows the keypoints correspondence to semantic information.
However, I am currently having some problems with 512x512 retraining, I hope someone can share their 512x512 training experience

[Qia98]

When I train the 512 model, I noticed that the vis picture seems to have been cropped

[Cdivyam]

This is the issue of imageio.imsave. It is slow to save the image through imageio.imsave. If you save the numpy file of the image and save convert it to image later then you can see there won't be any cropping.

You can also use cv2.imwrite which is faster and doesn't give the cropping error.