samyak0210 / vinet Goto Github PK

Hello, unfortunately link to model weights doesn't work. If you follow it, you will get the message 404 FILE NOT FOUND

Hi, I am having trouble identifying the steps necessary reproduce the results (AViNet only) for tables 4 & 5 in your Arxiv document. I was able to get extract the maps like the README.md describes, could not find the code to generate the metrics using those maps.

Datasets: DIEM, Coutrot1, Coutrot2, AVAD, ETMD, SumMe.
Metrics: CC, sAUC, AUC, NSS, SIM.

Could you please provide the steps necessary to replicate such experiments?

Thanks in advance!

Could you please show me the details of fine-tune for Hollywood-2 and UCF-Sports ? I want to know how you fine-tune the other datasets in code. And are use the same code of metric for this 3 datasets?

I was trying to evaluate a model after training. I noticed that they didn't release the ground truth labels of the test dataset.

In the evaluation code provided by https://mmcheng.net/videosal/

I found the comments are "if the ground truth cannot be found, e.g. testing data, the central gaussian will be taken as ground truth automatically."

However, the real code is:

if exist(saliency_path, 'file')
       I = double(imread(saliency_path))/255;
       allMetrics(i) = fh( result, I);
else      
       allMetrics(i) = nan;
end

Then in the end,

allMetrics(isnan(allMetrics)) = [];
meanMetric = mean(allMetrics);

I'm wondering for test set without ground truth, how to generate "central gaussian "

Another question is, for the numbers listed on the board https://mmcheng.net/videosal/, are they tested on validation set or test set?

Thanks a lot for your help!

Hi, I have some problem with the environments, the version of some packages doesn't match. Could you please provide some details about your environment requirements? Thank you very much.