Why Frame Crop only use the first frame?

[w3i1ong]

In the Dataset class AVLip, the crops are generated with image[:,0:500],image[:,1:501],...,image[:,4:504], and image[:, 504:2500] is not used. Is this a special design?

[JiaLiangShen]

I also find this bug, very strange.

[JiaLiangShen]

I also find this bug, very strange.

I think this is the proper way.

[FastWind123]

I think this is the proper way.

I have tested with this code on the val dataset, and the accuracy is 0.9368, which is nearly the same with the original bugged code. Very strange.

[142f]

I think this is the proper way.

I have tested with this code on the val dataset, 和 the accuracy is 0.9368, which is nearly the same with the original bugged code. Very strange.

I used python validate.py --real_list_path ./datasets/val/0_real --fake_list_path ./datasets/val/1_fake --ckpt ./checkpoints/ckpt.pt to validate on the val dataset. The validation results are: acc: 0.7812683499706401, ap: 0.7397901201278728, fpr: 0.3378778897451097, fnr: 0.10180337405468295. May I ask how you achieved an accuracy of 0.9368, and whether you encountered any errors that prevented the code from running properly during the code execution?

[FastWind123]

I think this is the proper way.

I have tested with this code on the val dataset, 和 the accuracy is 0.9368, which is nearly the same with the original bugged code. Very strange.

I used python validate.py --real_list_path ./datasets/val/0_real --fake_list_path ./datasets/val/1_fake --ckpt ./checkpoints/ckpt.pt to validate on the val dataset. The validation results are: acc: 0.7812683499706401, ap: 0.7397901201278728, fpr: 0.3378778897451097, fnr: 0.10180337405468295. May I ask how you achieved an accuracy of 0.9368, and whether you encountered any errors that prevented the code from running properly during the code execution?

This is weird. I just follow his instruction. I did not change the code.

[142f]

That's indeed quite strange. During testing, I encountered about 3 to 5 issues that prevented the code from running properly, and only after making modifications could it operate normally. Afterwards, I conducted tests on two different platforms: On Windows, the result was 0.953 (acc: 0.9530240751614797, ap: 0.9311341746150406, fpr: 0.048014226437462955, fnr: 0.04595695171611402); on Ubuntu, it was 0.9368 (acc: 0.9368761009982384, ap: 0.9214285883021583, fpr: 0.09662122110254891, fnr: 0.030250145433391506). These results match yours.

I think this is the proper way.

I have tested with this code on the val dataset, 和 the accuracy is 0.9368, which is nearly the same with the original bugged code. Very strange.

I used python validate.py --real_list_path ./datasets/val/0_real --fake_list_path ./datasets/val/1_fake --ckpt ./checkpoints/ckpt.pt to validate on the val dataset. The validation results are: acc: 0.7812683499706401, ap: 0.7397901201278728, fpr: 0.3378778897451097, fnr: 0.10180337405468295.5月I ask how you achieved an accuracy of 0.9368, 和 whether you encountered any errors that prevented the code from running properly during the code execution?

This is weird. I just follow his instruction. I did not change the code.