Roll training of head posture abnormal

[zlyxm]

1. 300w-lp-roll-krls_0.001_0.1.dat Training file size comparison ：

dataset 4000 Roll ：240MB， Pitch：14.9MB， Yaw：23.3MB

dataset 5000 Roll ：337MB， Pitch：18.4MB， Yaw：27.3MB

dataset 8000 Roll ：694MB， Pitch：28.1MB， Yaw：38,1MB

Obviously 300w-lp-roll-krls_0.001_0.1.dat is more than ten to twenty times larger than Pitch or Yaw

2. Roll testing num_wrong is too large：

Head Pose Training testing num_right: 872 testing num_wrong: 128 testing accuracy: 0.8720

and me

testing num_right: 951 testing num_wrong: 49 testing accuracy: 0.9510 train accuracy: 0.9998, test accuracy: 0.9510

But Pitch and Yaw testing accuracy are both greater than 0.96

3. The actual test roll error is too large

Pitch and Yaw are very easy to use, but Roll is wrong

I don't know where the problem is, can you help me?

[takuya-takeuchi]

@zlyxm Roll is very hard to detect exactly. So machine learning Krls could haave a hard time. This issue happens for me too.

[zlyxm]

yes,Maybe the data set is insufficient. Maybe too many key points cause interference. Generally speaking, when we recognize the roll of a face, we only need to judge whether the eyes are parallel, and whether the center point of the eyes and the chin form an inverted triangle, etc.

[veersamrat]

I need these data set all .dat files. could you help me ?

[aalexvi]

I need these data set all .dat files. could you help me ?

find you the files? i can´t find it

[takuya-takeuchi]

https://github.com/takuya-takeuchi/FaceRecognitionDotNet/tree/master/tools/HeadPoseTraining#2-download-train-and-test-data