search

D-X-Y / landmark-detection

Four landmark detection algorithms, implemented in PyTorch.
https://xuanyidong.com/assets/projects/TPAMI-2020-SRT.html
MIT License
925 stars 180 forks source link

Hi！ some problems about data process #49

Closed wqz960 closed 5 years ago

wqz960 commented 5 years ago

It is about the project of SAN. If I want to train the network using my own dataset, how to use PS to process the three kinds of pictures. I have not used the PS before, can you tell me how to handle the pictures and the setting parameters of PS?(gray, sketch, light). Thank you!!!

D-X-Y commented 5 years ago

I'm not familiar with PS. My dataset is created by PS with the assistance of my friend. He first creates the template for gray, sketch, light, and then applies then to all images of the original 300W and AFLW. Unfortunately, we have lost the original template. Sorry for the inconvenience.

wqz960 commented 5 years ago

@D-X-Y Thank you all the same for your quick respond! If you find the template some day, please contract me! Thank you! I am afraid that if you make the style dataset by my own template, the style-fusion module may not work!

wqz960 commented 5 years ago

@D-X-Y Sorry to bother you again! It is about the TS network!

`1.Predict HMi on both L and U using Eq. (5), and denote U with its pseudo labels as U1 . update the first student

  1. Optimize teacher with wg by minimizing Eq. (4) on L with prediction HiM and ground truth label Hi∗
  2. Compute the quality scalar of each sample in U1 using the optimized teacher via Eq. (3) 4.Pickup the top r × i × jUj samples from U1, named as L1ex
  3. Retrain w1 on L1 = L [ L1 ex by minimizing Eq. (2)` This is quoted from your paper what is the meaning of sentence 1? U1 is not labeled, how to update the first student. the effect of teacher network is let the student network pay more attention to the unlabeled images with low error and overlook the ones of high error?
D-X-Y commented 5 years ago

@D-X-Y Thank you all the same for your quick respond! If you find the template some day, please contract me! Thank you! I am afraid that if you make the style dataset by my own template, the style-fusion module may not work!

Thanks for your reply. Changing the template may cause SAN ineffective. You could post some examples and we can discuss.

D-X-Y commented 5 years ago

@D-X-Y Sorry to bother you again! It is about the TS network!

1.Predict HMi on both L and U using Eq. (5), and denote U with its pseudo labels as U1 . update the first student 2. Optimize teacher with wg by minimizing Eq. (4) on L with prediction HiM and ground truth label Hi∗ 3. Compute the quality scalar of each sample in U1 using the optimized teacher via Eq. (3) 4.Pickup the top r × i × jUj samples from U1, named as L1ex 5. Retrain w1 on L1 = L [ L1 ex by minimizing Eq. (2) This is quoted from your paper what is the meaning of sentence 1? U1 is not labeled, how to update the first student. the effect of teacher network is let the student network pay more attention to the unlabeled images with low error and overlook the ones of high error?

Thanks for your question. We use the predictions on U1 as pseudo labels, and use pseudo labeled data to re-train the student. The teacher network learns to select "good" pseudo labeled data for re-training.

D-X-Y commented 5 years ago

Feel free to open it if you have more questions. @wqz960