Custom dataset and points about face_landmark_factory HOT 9 OPEN

1. You can train 192 or 168 points - model with this set of tools, for more info, read the code.
2. You can train on other dataset, generally, after training 10,000 images, you could find the difference, training above 100,000 images, you could get ideal accuracy.

from face_landmark_factory.

@songhengyang THnaks for the response , the current network is working for img size of 64 , can it be changed to other image size. How will be accuracy affected

from face_landmark_factory.

Current network could work with image of other size. Image of large size will be more accurate, but the training speed will consume more time, images of small size will affect the accuracy, but the training speed will be faster.

from face_landmark_factory.

@songhengyang THanks for the response . Below are few queries

1. I want to train on 128x128 input image , where all should i make the changes in the code , i am using squeezenet
2. Can i train on grayscale 16bit input image which are annotated by me

from face_landmark_factory.

Suggest you check and try to modify the code by yourself, or you may wait, for I am planning to release a new version to solve these problems recently.

from face_landmark_factory.

@songhengyang sure i shall fork this repo and see if can make some changes . Just wanted to knw the tentative timeline for the next release

from face_landmark_factory.

We plan to release new version next month. App will train model on Ubuntu 18 and 16, not windows.

from face_landmark_factory.

@songhengyang thanks for the update and response . If you could give me some pointers on changing the image size during training it would be helpful

from face_landmark_factory.

check and change parameters in script directly.

from face_landmark_factory.