Closed ebraheemsaleh closed 3 years ago
Actually face images and marks.
I am using face-mesh-generator's features/tfrecord-marks-68 branch to generate dataset: 300vw, 300w, AFLW2000_3D, afw, helen, lfpw as train.record ibug as val.record Training 1epochs with lr=0.01, 10epochs with lr=0.001, 100epochs with lr=0.0001 After training complete, I also transfer model to tensorflow frozen model (.pb only) What I found is that the prediction of mouth is very good. But the prediction of eye is very worse. The prediction result of the eye always open, even I close my eye. Any suggestion?
It is time to move on to a more advanced project -> https://github.com/yinguobing/facial-landmark-detection-hrnet
If you would like to stay with this one, I recommand starting with weighted loss function.
Please open a new issue next time. I think this issue had gone stale and better to be closed.
Which dataset train.record should point to? Dataset of box annotated faces ?