Open waHAHJIAHAO opened 1 month ago
I am not sure if we experienced the same problem. I think this is something you should ask to the original authors of the face landmark detector. Perhaps, they correspond to the frames were no face was found, but I don't think so, because we were handling these situation by zero-filling the feature sequence. If these NaN frames are no so common, I would recommend you replace them by zeros, using this code:
face_landmarks[np.isnan(face_landmarks)] = 0.
you can use do it in advance and saving again the files, or you can dynamically apply this logic when defining your Dataset, the object in charge of loading the data for training and evaluation.
I use private dataset which is collected from mental colleage in my universiy,.I processed my original video data in the way that D-vlog processed the data. I finished the data processing of landmarks and stored the corresponding npz file in the faces directory. When I ran the emonet script, there was a value of NAN in the processed result. May I ask why