Thank you for your interesting work!
I am a researcher from the TFace team. We are committed to reproducing the results in the paper. We have used the ’dcface_3x3.ckpt‘ to sample a synthesis dataset. The recognition model trained by this dataset achieved an average performance of 86.14% on five verification datasets. This is very similar to the 85.79% reported in your paper, but we found that the dataset generated in this way only performs 73.43% on CFP-FP. We see that you did not provide this result in your paper. Is this consistent with your experimental results?
In addition, when we replace it with ’dcface_5x5.ckpt‘, the performance of the recognition model suddenly drops to 50%, just like the setting of ’7x7‘ in your paper. We found that the results generated by ’dcface_5x5.ckpt‘ do not seem to maintain an identity. Is there anything wrong with ’dcface_5x5.ckpt‘? We only replaced the path of ckpt without making any additional modifications.
Thank you for your interesting work!
I am a researcher from the TFace team. We are committed to reproducing the results in the paper. We have used the ’dcface_3x3.ckpt‘ to sample a synthesis dataset. The recognition model trained by this dataset achieved an average performance of 86.14% on five verification datasets. This is very similar to the 85.79% reported in your paper, but we found that the dataset generated in this way only performs 73.43% on CFP-FP. We see that you did not provide this result in your paper. Is this consistent with your experimental results?
In addition, when we replace it with ’dcface_5x5.ckpt‘, the performance of the recognition model suddenly drops to 50%, just like the setting of ’7x7‘ in your paper. We found that the results generated by ’dcface_5x5.ckpt‘ do not seem to maintain an identity. Is there anything wrong with ’dcface_5x5.ckpt‘? We only replaced the path of ckpt without making any additional modifications.