tf-blazeface icon indicating copy to clipboard operation
tf-blazeface copied to clipboard

Low accuracy in inference

Open rose-jinyang opened this issue 5 years ago • 3 comments

Hello How are you? I've trained a model for detecting 6 keypoints and face bounding box. I used your project without any modification. I tested the model on some images but it's accuracy is very poor.

image

image

Even for some images, it is impossible to detect the face bounding box. Please let me know the reason. Thanks

rose-jinyang avatar Jun 23 '20 01:06 rose-jinyang

Hi, we can simply say that your model couldn't generalize very well. the300w_lp dataset is not enough to train a robust model by self. I just added it to create an entry point for those who want to work on this subject. As I said before, I do not intend to share the dataset I use at this stage. If there is a change in this matter, I will already make the necessary updates in the readme.

FurkanOM avatar Jun 23 '20 06:06 FurkanOM

Hi I think that the accuracy of model is very poor although the300w_lp dataset is not enough to train a robust model.

rose-jinyang avatar Jun 23 '20 09:06 rose-jinyang

Hi, Yes, as I explained above, your model does not have a generalization capability. I mean, the model you train is probably overfitted the300w_lp dataset.

FurkanOM avatar Jun 23 '20 10:06 FurkanOM