Train procedure image, bbox and labels

[DavideCatto]

I have some troubles during the training of RFB with the dimension of 320x240. In particular I'm following all the steps for train on widerface (I have downloaded VOC dataset from the link in description). Now I can t train and seem that the network doesn t learn: the loss remain stable and doesn t descrease. The imagas in input have a dimension of BxCxHxW and with value of -1.0 to 1.0 and the same augmentation of datasetVoc class. The labels in input have a dimension of BxN where N is 2. (Number of class). 0:background 1:face The bboxs in input have a dimension of BxP where P is the number of Priors. What I can t understand is how bbox/priors are calculated: I'm using "MatchPrior" to trasform bbox in the correct form but I don t know if it s the right things to do.

Can you please help me with some example transformed for image,bbox/priors and label to feed the network?