Open Svardfox opened 1 year ago
It seems there are overfitting problems..
For SAMM dataset, we crop and align the face based on the original data, you may try dlib toolbox or some other methods to preprocess the data and then do the traning.
Since face alignment and cropping are an important part of MMnet algorithm, could you please share detailed face alignment and cropping methods so that we can better understand your work?
I tried to reproduce the results of the network in the paper on the SAMM dataset, but I couldn’t get good results. The validation accuracy of most sub-networks is much lower than that of the training set, so I wonder what might be causing this?