I suppose it is the coordinates (x, y), but if you could tell me how to scale it to normal coordinates.
@hugochan

I am trying to run this in cmd but there is an error

python itracker_adv.py --train -i input_data -sm saved_model

how I can give the path of the dataset

def load_data(file):
npzfile = np.load(file)

thank you

开始以为是坐标，结果发现有正负···知道字段的含义吗？
[[ 1.70154930e+00 -1.93875000e+00]
[-6.85229867e+00 7.72708985e-01]
[-1.08925000e+01 5.00000000e-04]
...
[-7.52597535e+00 9.69375000e-01]
[ 5.78865727e+00 -4.60561372e-01]
[-3.64302465e+00 -9.69375000e-01]]

Hello dear author, where can I find this npz file，This website with http://hugochan.net/download/eye_tracker_train_and_val.npz seems to be inaccessible

Have been using Tensorflow 2.0 but I realized i had multiple issues calling functions like tf.session()

Hi, really appreciate if you could provide and share the python code to extract face, left eye, right eye and face grid from multiple images (or video / sequence of image frame). In short image.jpg to face.npy, LeftEye.npy, RightEye.npy and FaceGrids.npy. (or .npz file)

Hi Hugo Chan,

I'm a research staff based in Singapore, Nanyang Tech University.

I came across your version of iTracker am I'm really interested to study it and studying if I can implement it in the computer games I'm working on.

In the original paper of Eye Tracker for Everyone, it was mentioned that the number of subjects in the training set matters a lot to achieve generalization of the neural network. May I check with you while processing a smaller subset of the original data, how many subjects were included (Out of the original 1437 subjects) how many instances per subject?

Does the validation set share similar subjects and instances per subjects as the training set or does the validation set only include subjects not seen by the model when being trained by the training set?

Possible to share your code how you processed the original data set? :)

That aside,
Currently, I made an inference engine using openCV library but unfortunately using your pre-trained model as the deployed model, and my face through my laptop's webcam as the Inference instance, accuracy is horrible. The pre-trained model 23 was not able to detect my gaze at all. It gave random values.

What do you think may be the problem?

Guang Way

Hello. We are running this code in PyCharm. I know that there are 4 files that are created when saving in tensorflow. To train the model, we tried running the program by putting

"--train -i pretrained-models/itracker_adv/model-23.data-00000-of-00001 -sm saved model" in the paremeter value of the configuration, but the following error occurs.

If you know how to solve it and answer me, I would be grateful.
Thanks for reading my answer. Have a nice day!

Hi Hugo, I'm interested in studying your work for fun. I wanna first try to run it but I found no loaded model could be passed into the program. Could you please show me how to figure it out? Thx.

Hi,
Can you explain how to use the trained model to predict?

Thank you for your work. Could you please share a newer link for the smaller dataset, because the existing link seems to be broken. Thank you for your time

Hi Chan,
Can you please guide me, how to use dropout in Eye-Tracker net?

请问如何使用gazecapture数据集训练这个模型呢，这个数据集不是.npz文件

Hi Hugo. I see you have mentioned the small dataset in the readme. Can you provide me more information about the datasets? I would like to know how you obtain it.

Hi!

Do you know a good method for capturing the face parts from a video or a image?

Thanks

@hugochan hi thanks for the wonderful code for eye tracking and detection , just wanted to know whats the procedure to train on a custom dataset? can you briefly give the pointers for training on the new/ custom dataset

你好，请问train_y.npy、val_y.npy文件里只是单纯的（x,y）坐标，还是包含了前面四个图像输入的对应关系？求大神解答

Hi, I remember reading this paper some time back, and as I understand the device type and orientation mattered a lot. This is because the position of the camera relative to the screen makes a difference.

In this code, I don't see any parameters to specify device type or orientation. Wondering if you could please explain what I'm missing? Does this model only work for a specific device type and orientation? Which device and orientation would that be?

Thanks in advance!

Hello, I dont have an ML background, but would like to try this out.

How do I get the x and y gaze prediction output, given a face image, width and height of screen and maybe some camera details (distance of camera to screen, and camera size maybe)?

An example python implementation would be great, thanks a lot :D

Cheers!

P.S. Can this be used to for different screen size? like instead of an iPhone screen, a 24 inch monitor is used?

I expect both train_y and val_y are the 2D coordination of eye graze .

But the result is quite strange. Most of the coordination are very small or negative. That means most eye graze always point at the left top side. But the fact doesn't like that. Most of them are wrong.

Please concern the circle and the center of circle. The center is the predicted point of eye gaze.

Hi, Hugo Chan, nice to meet you, I am recently focused on this technology which might help us to get to know more about our end-user's thoughts, I read about the MIT paper, they claim their prediction error is 1.71cm on mobile and 2.53cm on a tablet.

And in this repo's README, it says" better validation error (2.19 cm vs. 2.514 cm)", is this the comparison to the MIT <eye_tracking_for_everyone> result?

Thanks!