dkollias / aff-wild-models Goto Github PK

Hi,
Thank you for sharing your work.
Sorry, I am new to deep learning. Could you show how to use the pretrained model to predict emotion from an image with vgg-face model and get the valence and arousal value from tensor with code or example?

Thank you for your time,
tsly

Hi,
I am new to deep learning. And I tried to run the model in tensorflow 2.0, but some functions, such as tf.contrib.slim cannot work any more. How can I solve it?
Thanks

Hi Dimitrios,

I'm trying to reproduce the results you got on the AFEW-VA database. Since the models shared are not fine-tuned for AFEW-VA, I think I should train it myself. Can you provide the script and the parameters you used for fine-tuning on AFEW-VA?

A second question is about the input images (faces). I use MTCNN as a face detector, did you use the provided landmarks to determine bounding boxes?

Hello dkollias,
Thanks for sharing your work! I'm interested in this data set and want to train a model by training all the faces which I crop based on bounding boxes. There are some problems I have:
1. I found that many faces I cropped from video are in the trailer like Jon Snow's face. Is the annotation of that face?
2. I want to train my model not using RNN or any sequence based methods. Have you ever tried to train a CNN to extract features and then feed that features to a regression method before? In paper, you just mentioned using SVR but not using a CNN model with it.
3. Can I get the test set annotation to evaluate my model? Since there is no test annotation, so I want to ask how to evaluate my own data set without test set ground truth. To evaluate my models before, I analyzed test accuracy by comparing ground truth and my predictions.
This is my first time to use a huge data set and train a non-classification model. I appreciate your patience to read and help.

Thanks,
Andy

Hi,

Thank you for providing the pretrained weights and evaluation codes. I tested the vggface model on the Rafd dataset and found that the va values of some categories such as sad or fearful vary a lot. They may even cross the axis. I only used the frontal face and cropped them before evaluation. Because the Rafd dataset was collected under lab environment, I thought each category should span a constrained region in va plane. It turned out not. Do you know any reasons about this? One cause I could guess is that your models were trained with wild dataset which may not fit in the lab controlled data.

Best,
Li

Here is the scatter plot of va points per each emotion in Rafd dataset:

The codebase uses deprecated tensorflow calls such as tensorflow.contrib
Can a requirements.txt be added, to know which versions of the supporting library we need?

Hello Dimitrios,

From what I understand in the paper, the model would take both face image and its 2D landmarks as input for training the network, could you please let me know, how did you train your model on both image and 2D landmarks ? did you inject it somewhere in the FC layer or concatenate with the image ?

Thanks,