can't find computer_vision_utils file

[NingMingHao]

Recently I'm trying to achieve your experiment, but I can't find the computer_vision_utils file, so I self-defined the read_image function as follows:

def read_image(file_path,resize_dim,channels_first=True,color=True):
    if color:
        raw_image = cv2.imread(file_path)
    else:
        raw_image = cv2.imread(file_path,0)
    raw_image = raw_image.astype(np.float32)
    resized_image = cv2.resize(raw_image,resize_dim)
    if channels_first:
        return resized_image.transpose(2,0,1)
    else:
        return resized_image

But the prediction performance is really bad, so I'm wondering if I have missed some important operation? By the way, is the 'dreyeve_mean_frame.png' generated as the mean of the first 37 runs? And the semseg_branch also performs badly, I think there must be some wrong with my code. So I attach my code here, please help me. Here is my models.py: models.py.txt

[ndrplz]

Hi @NingMingHao ,

I can't find the computer_vision_utils file, so I self-defined the read_image

read_image refers to this implementation. Probably we should add this kind of utils functions inside the dreyeve repo, thanks for pointing it out.

is the 'dreyeve_mean_frame.png' generated as the mean of the first 37 runs?

It is indeed.

semseg_branch also performs badly

I don't understand if the problem lies in the semantic segmentation network or in the dreyeve branch that predicts the gaze from the segmentation. Anyway, two suggestions:

• Use a more recent pretrained network for segmentation, such as the mapillary model used here.
• Remove the segmentation branch to sanity-check that the rest of your model works properly. From ablation study in the paper (Table 4) you can notice that the segmentation branch is the one that contributes the less to the final result; you can safely disable it to see if things still work.

Best, A

[NingMingHao]

Really thanks for your rapid reply. And I will try your suggestion to remove the segmentation branch. The reason why I upload my models.py is that I'm using tensorflow as the backend of keras, it's a pity that tensorflow doesn't support to resize a tensor using scale_ratio, so you can find that I have commented some your code, and used coarse_h = Lambda(lambda x: tf.transpose(tf.image.resize_bilinear(tf.transpose(x,perm=[0,2,3,1]),[_w*4,_w*4],name='{}_4x_upsampling'.format(branch)),perm=[0,3,1,2]))(coarse_h) Maybe there is something wrong here, and I will check it out. Thanks again!

[NingMingHao]

I have test this code: coarse_h = Lambda(lambda x: tf.transpose(tf.image.resize_bilinear(tf.transpose(x,perm=[0,2,3,1]),[_w*4,_w*4],name='{}_4x_upsampling'.format(branch)),perm=[0,3,1,2]))(coarse_h) I'm sure it works properly,

1. 8x8 resize in

2. 8x8 resize out

3. 4x4 resize in

4. 4x4 resize out

And finally, I have a look at your model weights, I find that all the convolutional bias weights of SaliencyBranch are 0, I'm not sure if this is the reason?

finally, this is the output the im_net gives.

[varunjammula]

@ndrplz Hi, I am trying to run predict_dreyeve_sequence.py file. It requires dreyeve_mean_frame.png file. How can I generate this file?

[ndrplz]

Hi @varunjammula you can generate it as the average of all frames of all sequences of the training set.

[varunjammula]

Thanks for clarifying the issue above. I have a new issue now. Where can I get dreyevenet_model_central_crop.h5 model?