The performance of the pretrained model you provided is somewhat different

[tfygg]

Hi, i tested the pretrained model (300W-LP, alpha 2) you provided on AFLW2000, and found that the performance is somewhat different from shown on the paper? is this model the best of your algorithm?

[natanielruiz]

Hello tfygg,
Please keep in mind that since our method predicts 66 bins of 3 degrees each - the minimum and maximum degree angle that it can predict is -99 and +99. For that reason we filter our test datasets and remove images of angles higher than that. We use 1969 of the 2000 AFLW2000 images.

I just re-ran the test on those images and I have the reported accuracies.

[natanielruiz]

[natanielruiz]

Oh by the way I can send you the image list that I use for AFLW2000 by email if you want 🙂

[natanielruiz]

Please tell me if this fixed the mismatch, also check if you're using the test_hopenet.py script to test it. Thank you!

[tfygg]

Ok, thanks. My email is tfy7@qq.com.
By the way, When trained the model, should i filter the datasets and remove the images?
Can you send the train list to me，please?

[tfygg]

Thank for your help.
I used the image list that your provided, the accuracies is no problem.

[tangbohu]

Dear Natanielruiz, could you also send me the image list?
My email is tangbohutbh@gmail.com. Thanks a lot!

[fengziyong]

@tfygg @natanielruiz
I filtered the 300W_LP and AFLW2000 by myself. But I can't reproduce result reported by the paper.

I train on 300W_LP and test on AFLW200.
Settings:
python train_hopenet.py --output_string=2018-06-09_1 --data_dir=../data --filename_list=../pose_clean.list --alpha=1 --lr=0.00001 --gpu 3 --num_epochs 25

[shz0519]

Dear @natanielruiz @tfygg @tangbohu ,could you send me the 1969 of the AFLW2000 image list?
My email is shz_0519@126.com. Thanks !

[Zouhj]

Thank you for your email. I'll disposal and reply as soon as possible.Best Regards!邹涵江