danims-zgz / panoramic-cnn-360-saliency Goto Github PK

Hello,

I'm trying to reproduce the scores by using this code.
But the prediction was still far from the score on the paper.
Could you share the best model file that we can inference from it?

Many thanks.

Hi,

Thanks for your kindly sharing of your code.
However, there's no data augmentation code mentioned in paper. (7 data augmentation)
Could you also share the pre-processing code here?
By the way, I've tried to reproduce the score in paper.
But the output color seems to be opposite. (black to white, white to black)
Is there any post-processing steps to the output from the model?

Many thanks.

gt

my reproduce result

Hi, it's me again.

I tried the evaluation code that you shared, and changed the ground truth to the head movement only heatmap. (Like bellow)
(read them as grayscal)

The similarity and cc score are the same as you mentioned last time.
But the scores calculated with binary fixation maps are still different from the scores you gave, especially the NSS score.
I parsed the binary groundtruth by the code released in salient!360 2017 with 'SP*.txt' files. (code)
I doubted that we did the same in this parsing step.
Could you also share the parsing code of binary fixation map?
Many thanks.

The scores are as follows.

By the way, it's also strange that the KLD score seems to be better than that you mentioned.