Request additional information for paper replication about faldetector

[PeterWang512]

1. https://drive.google.com/file/d/1mzNxCyrUTBF7-lQGPLYT0HuUODvVvtsb/view
   Training set and validation set will not be released due to licensing issue
2. #3
3. For each paired real and fake image, compare the model output of the two images. If the model's output indicates the fake image is more fake than the real one, then this serves as a correct count for 2AFC. You can then take (total correct count) / (total pairs) to get 2AFC
4. EPE: for each pixel, take the Euclidean distance between the GT and predicted flows, and average across the pixels to get EPE for a single sample. Then, average across samples to get the metric.
   IOU-3: take the magnitude of the flow for each pixel, then generate a binary mask using 3 as the threshold. Calculate IOU between the mask version of GT and predicted flow for each sample, and average the IOUs across samples.

Hope this helps, and for more information please see the paper: https://arxiv.org/pdf/1906.05856.pdf

[TaeniKim]

Thank you for your quicky reply..^^