Open kkkmax opened 5 years ago
Later, I measured my own picture and found that I need to process the picture into a 6-channel picture. I see that the 4th and 5th channels of the information you added are dependent on the structured information of data/tracker/00xxx.mat. Read out is a 49*2 list information, but I don't know how to generate this information. Can you provide a way to generate it? This will help my research greatly, if you can, thank you very much! !
In addition, I saw that someone used the information of the first 49 key points of the face 68 key point to replace the "data/tracker/00xxx.mat" file. I used this process to try and perform very badly.
The picture order correspondence is consistent with the above, I just kept the information of 1 channel, the error is significant.
The way I tried above, I think it can only be performed on a portrait with a clear face. When I tested it, the test gave a photo of a dog. He tested it as a human figure! !
I would also like to ask what is the requirement for the picture when testing, what is the ability and deficiency of the model prediction?