Some question about dataset in the paper

[dannyyangya]

Thanks for the perfect work.

I'm confused by some dataset settings in the experiment from the paper.

Traffic-Train Dataset: It was mentioned in the paper that 4160 frames were used for testing in your experiment, which was not aligned to the source dataset from http://vision.eecs.yorku.ca/research/anomalous-behaviour-data/. Could you help to provide the testing setting of your experiment on this dataset, i.e. how to get the testing set of your experiment from the source testing set provided by Andrei Zaharescu and Richard P. Wildes?

Subway Dataset: After a survey on this dataset, I found that the annotations were redefined. And I found the redefined annotations from http://vision.cs.utexas.edu/projects/abnormalactivity/cvpr2009-annotation.htm. Is this the same definition for your experiments?

Thanks!

[nguyetn89]

Hi @dannyyangya,

Our experiment in Traffic-Train was different from the dataset's source because we found some difficulties when following its description. That is why we re-implemented some other methods and performed the experiments in another way for comparison. For the Traffic-Train, first 800 frames were used for training and 13840th-18000th frames were for evaluation. You may take a look here.

For the subway dataset, we used the annotation in the link you mentioned (maybe with a little bit modification in training part, I am not sure). You may check the annotation here. That repository is for another improvement of this work, but it should have a similar configuration.

Hope this helps!