Code release date

[leopqrz]

Hello, I know you guys said the code will be released soon but, I'm wondering if there's a close date I can wait for. I need a good face landmark detection for my application and I believe this is the SOTA. Thank you.

[maxcrous]

Not one of the authors, but in the mean time you could get very near SOTA by using stacked hourglass networks like Face alignment networks. There is a pytorch implementation here, or pm me for a tensorflow rewrite.

[leopqrz]

Hi Max, thank you very much, I've already implemented it and it's very good. But it didn't work for some of my images. Retinaface was better but it’s only 5 landmarks. Do you think is it better than HRNet?

On Sat, Feb 20, 2021 at 1:46 AM Max notifications@github.com wrote:

Not one of the authors, but in the mean time you could get very near SOTA by using stacked hourglass networks like Face alignment networks https://arxiv.org/abs/1703.07332. There is a pytorch implementation here https://github.com/lippman1125/pytorch_FAN, or pm me for a tensorflow rewrite.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/nothingelse92/AnchorFace/issues/2#issuecomment-782589065, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADRYAACFDONE64RIOLQL4O3S75ZHNANCNFSM4X5Q3RLQ .

[maxcrous]

Nice that you got it working 👍 I wouldn’t know, I haven’t read the paper or seen it in a comparison. Landmark detection is nearly “solved” in the sense that the mistakes these models make are often due to bad annotations (at least, that’s what the FAN authors suppose). So it could be a data problem, i.e., your target data differing from the model’s training datasets. One approach could be to retrain and use aggressive augmentations.

[gofixyourself]

@leopqrz if you need an accurate and fast model for facial landmarks then try this

[hzxie99]

@leopqrz Hi, if it is convenient for you, could you tell me what method you use? Because the 5 landmarks don't fit my images either.