Cho-Ying Wu

image resolution and bbox cropping may affect the performance. You can try to tune the detected bbox size (increase or reduce the size a little bit) for the best performance.

We chose the best checkpoint among all epochs. You can save more checkpoints during training and evaluate them to find the best one.

The AFLW2000-3D are cropped based on 3DDFA (https://github.com/cleardusk/3DDFA/tree/master/test.configs). This benchmark script is for validating our performance. We do include a face detector for processing in-the-wild images and will release that...

Oops. I uploaded the wrong model. I've updated the readme and the link. Please check.

aflw2000-3D is shared in the link (ReadME, Single Image Inference Demo - Step4, Download the data) python benchmark.py -w "pathToPoseModel", you'll get the reported number.

Thanks for your sharing. In the first example, the performance should be optimal if the face is center-cropped. We'll release the single-image inference demonstration recently.

The full codes are released.

yes, its pose, shape, expr for 12:40:10

We will soon give an update on the other part code release.

The training codes are released. We'll release the rest part of codes this week.