Questions related to the bbox and landmark values about img2pose HOT 3 CLOSED

Hello Sungjun Ethan Yoon,

Thanks for your interest in our work!

Regarding your questions:

1. As RetinaFace provides manually annotated landmarks for only 76k training faces and none for validation faces, we use its prediction for the rest of the training faces (63k) and validation faces (34k) that do not contain landmarks. For the manually annotated landmarks, the original bbox is used to estimate the 6DoF. For the other ones, we use the bbox prediction, as in our experiments it resulted in better 6DoF estimation since the predicted landmarks are more centered.

2. For some faces, we refined the 5 landmark predictions from RetinaFace using a 68 landmark detector by matching the center of the eyes of the two landmark sets. But this step is not necessary, as similar results are achieved while only using RetinaFace's 5 landmark predictions.

3. For training, it won't matter which bbox is stored, as the bbox is actually projected from the 6DoF during training, and is customizable. And, as I said in answer 1, we used a mix of WIDER FACE GT bbox and RetinaFace bbox to estimate 6DoF labels for training.
   The original and img2pose bbox you posted are in different formats: the original is left top width height; while ours are left top right bottom.

Hope this answers your questions.

from img2pose.

Dear Vítor Albiero,

Thanks for helpful answers.
I clearly understand what you are saying, and all of questions are solved due to your answers.

Could I get the information about the used 68 landmark detector (paper or code)?

Best,
vujadeyoon

from img2pose.

Yes, we used this code.
It has good predictions when dealing with large faces, but it is very unstable with medium and small faces (which WIDER FACE is mainly composed of).

from img2pose.