search

1adrianb / face-alignment

:fire: 2D and 3D Face alignment library build using pytorch
https://www.adrianbulat.com
BSD 3-Clause "New" or "Revised" License
7.05k stars 1.34k forks source link

Comparison with the improved 3DDFA #122

Closed segalinc closed 5 years ago

segalinc commented 5 years ago

HI, thank you for your contribution on this topic. I am curious to understand how your approach compare w.r.t the new improved version of 3DDFA reported at https://arxiv.org/pdf/1804.01005.pdf

Thank you.

1adrianb commented 5 years ago

Hi @cristinasegalin, while the authors of 3DDFA do not offer a direct comparison, judging from the results reported in [https://arxiv.org/pdf/1804.01005.pdf, Table 3] and the ones from [https://ibug.doc.ic.ac.uk/media/uploads/documents/trigeorgis2016mnemonic.pdf (MDM), Table 1] we can notice that new 3DDFA has approximatevely the same accuracy as CFSS on 300W. Since MDM significantly outperforms CFSS and our method surpasses MDM, most likely on average our method will outperform the improved 3DDFA results by a helthy margin. In addition, our method doesn't suffer from the failure cases reported in Fig. 20. That beign said, while this is the most likely conclusion, the best will be to evaluate them on a common dataset, ie. 300W or LS3D-W-Balanced.

segalinc commented 5 years ago

Thanks for the quick reply. Really appreciated.