some other networks do not converge in the second stage.

[wqz960]

Hi! Thank you for your excellent work, I just replaced the network which is vgg in your original network with Resnet or some others for my interest. during the first-stage training, the network shows almost the same performance with vgg. But in the second-stage training. it seems that the network no longer converges, and the effect is even worse than that in the first stage. According to the concept of heatmaps and feature maps in your paper, there should be some improvement in the accuracy in the second stage, but during my training in the second stage, the validation accuracy has decreased. By the way, in the second stage, I deleted the face alignment operation, the connected section only contains the heatmaps and feature map process second. whether this is the cause of the problem? In my opinion, it does not matter. Looking forward for your responds!!!

Thanks!!!

[MarekKowalski]

Hi,

Do you mean that the second stage no longer gets a face image as input? If that's the case then it probably won't work.
You can provide the second stage with a non-aligned image, essentially the same as intput to the first stage. However, if you do that, remember that the heatmap should also be generated from landmarks that have not been processed by the connection layers.

If that's what you are doing then it should work fine.

Best regards,

Marek

[wqz960]

hi! Thank you for your quick respond. The feed to the second stage is a cropped face image which is not aligned, the feature map and the heat maps generate from the non-aligned landmarks . But in that case the network doesn’t converge. By the way, the input size, I have modified to 2562563, does the different input sizes cause the problem? can you provide me some advice to solve this situation? Thank you!

[wqz960]

sorry to take your time again! I saw some other network including yours, when the images change from the original size to 112 size, the groundtrue landmarks should minus 1 because of the affine tranforms. when should I minus 1 during training? when changing the original size to 256 size, whether the groundtrue need to minus 1?